Meeting Title: Placeholder | CTA Working Session Date: 2026-05-06 Meeting participants: Amber Lin, Katherine Bayless, Uttam Kumaran
WEBVTT
1 00:00:15.620 ⇒ 00:00:16.520 Katherine Bayless: Whoa!
2 00:00:19.980 ⇒ 00:00:20.950 Amber Lin: Heather!
3 00:00:21.410 ⇒ 00:00:36.739 Katherine Bayless: Well, hello. I totally, I was on a different call before this, and I was like, I’ll just have to drop halfway through, no big deal, and I totally did the that guy thing, and didn’t pay enough attention, because I didn’t realize I was the host as I end meeting for all, and I was like, well, sorry guys.
4 00:00:39.540 ⇒ 00:00:44.850 Katherine Bayless: It’s like the, the worst version of, like, you’re on mute. Like, come on, it’s 2026, Catherine, we know better.
5 00:00:44.850 ⇒ 00:00:45.630 Amber Lin: But…
6 00:00:47.370 ⇒ 00:00:48.440 Katherine Bayless: Let’s see…
7 00:00:48.830 ⇒ 00:00:58.930 Amber Lin: All good, I think Item should have time… now. I know he was also in another meeting, so I think we can get started, and he’ll probably drop by.
8 00:00:59.210 ⇒ 00:01:00.480 Katherine Bayless: Okay, that works.
9 00:01:01.360 ⇒ 00:01:02.410 Katherine Bayless: Yeah.
10 00:01:02.410 ⇒ 00:01:04.300 Amber Lin: Hi, Tom! Oh, he’s connecting.
11 00:01:04.410 ⇒ 00:01:05.500 Amber Lin: Bye, Tom!
12 00:01:06.320 ⇒ 00:01:07.410 Uttam Kumaran: Hello!
13 00:01:07.730 ⇒ 00:01:09.310 Katherine Bayless: Hello, hello.
14 00:01:09.970 ⇒ 00:01:15.500 Uttam Kumaran: I had a great call with Jay this morning, actually. We have some new, like, thoughts on…
15 00:01:15.690 ⇒ 00:01:20.179 Uttam Kumaran: Like, context engineering, and we had a good, sort of, Morning, a little bit on…
16 00:01:20.670 ⇒ 00:01:26.980 Uttam Kumaran: All things AI, so I think we’ll try to reconvene on some stuff next week when we’re there.
17 00:01:27.690 ⇒ 00:01:28.870 Katherine Bayless: Nice, yeah.
18 00:01:28.870 ⇒ 00:01:53.449 Katherine Bayless: I mean, yeah, so, as, I’m sure he talked about it, because he went straight from the meeting that we were in to talk to you, and then, that I just accidentally ended for everybody. But, yeah, so we talked to the two, like, my team, his team, this morning about the, the speaker selection, stuff, which I think I probably mentioned, like, in my normal, chaotic, not really enough information sort of way on Monday.
19 00:01:53.500 ⇒ 00:01:59.840 Katherine Bayless: But basically, the TLDR version is that the conferences team has…
20 00:02:00.130 ⇒ 00:02:03.100 Katherine Bayless: Chosen us to build their software this year.
21 00:02:03.100 ⇒ 00:02:03.510 Uttam Kumaran: Yes.
22 00:02:03.510 ⇒ 00:02:10.580 Katherine Bayless: So we’re going to design in-house the session management, the speaker submission slash application, the
23 00:02:11.039 ⇒ 00:02:17.429 Katherine Bayless: all of that. And, like, I think, you know, to me, the… like…
24 00:02:17.429 ⇒ 00:02:42.349 Katherine Bayless: the real value here is that basically this is the first baby step towards, like, everything comes through, like, our CRM, essentially, is what we’re gonna one-night building. It’s gonna be a massive mess of microservices probably for a while, but, like, we’re starting to actually control where the data, like, is created, and where it lives, and how it moves, and what it looks like, and so, like, my team won’t be needing to ingest data from a ton of different platforms, though there will still be some.
25 00:02:42.349 ⇒ 00:02:56.409 Katherine Bayless: Right? But, like, all of the first-party stuff will actually be first-party instead of, like, vendor-supported. So it’ll take time to get there, but I think if we can, like, nail this first one, the buy-in will be swift.
26 00:02:56.409 ⇒ 00:03:05.609 Katherine Bayless: And probably slightly overwhelming. But yeah, so, it’s like all eyes on being able to deliver and govern data at scale.
27 00:03:06.420 ⇒ 00:03:16.159 Uttam Kumaran: Great, I mean, it’s kind of… yeah, it’s exactly sort of what I mentioned to Jay, which is, like, we have to ingest more data than was… we would have taken just for reporting use cases.
28 00:03:16.770 ⇒ 00:03:32.999 Uttam Kumaran: And you have to manage, sort of, like, any integration surface, so, like, relational database may be one. We may need to hook into, like, telemetry or logs somewhere else. We also have all the MCPs, CLIs, and then, not only are you enabling chat interface, you’re enabling
29 00:03:33.040 ⇒ 00:03:41.999 Uttam Kumaran: applications. So, some of that may be traditional apps, some of that may have agentic use cases, but all of that sort of is powered by the same
30 00:03:42.100 ⇒ 00:03:45.689 Uttam Kumaran: really good backend database, basically.
31 00:03:46.070 ⇒ 00:03:59.540 Uttam Kumaran: And then, really, another point that I told him was, like, look, I think you’re gonna have a situation, because we were talking also about the people, and I think there’s gonna be, even in our company, there’s just gonna be some people that are creating skills, using skills.
32 00:03:59.540 ⇒ 00:04:09.109 Uttam Kumaran: But they’re really, like, training others or partnering with the subject matter experts to, like, understand what software needs to get built. And then there’s, like, a small core of people that are, like.
33 00:04:09.180 ⇒ 00:04:24.009 Uttam Kumaran: actually building, like, doing the hardcore engineering of building all of that, you know? And I think those are, like, two separate profiles, and I think it’s actually much weighted towards the people that are, like, out in the field understanding the requirements.
34 00:04:24.100 ⇒ 00:04:30.170 Uttam Kumaran: Like, working with the subject matter expert, doing a proof of concept of something, and then sort of, like, handing it to, like.
35 00:04:30.620 ⇒ 00:04:33.749 Uttam Kumaran: Kind of like a software factory-style team, you know?
36 00:04:34.130 ⇒ 00:04:44.819 Katherine Bayless: Yeah, I mean, it’s funny, right? Like, it’s like, at the end of the day, probably the familiar pattern will persist from, you know, user has idea, user scopes idea, user designs what they’re looking for.
37 00:04:44.820 ⇒ 00:04:45.270 Uttam Kumaran: Yeah.
38 00:04:45.270 ⇒ 00:04:53.859 Katherine Bayless: describes requirements, and then, you know, technical team helps execute. It’s just, to your point, like, I think people will be able to come further down the line than they used.
39 00:04:53.860 ⇒ 00:04:54.510 Uttam Kumaran: Yes.
40 00:04:54.680 ⇒ 00:05:01.980 Katherine Bayless: like, I think we’ll be able to expand the number of people who can be the technical folks. Like, if you’re just building
41 00:05:01.980 ⇒ 00:05:18.139 Katherine Bayless: So, like, the VP for membership, she needs an application that will basically be web form reviewers and, like, you know, system of record on decisions having been made for innovation ideas to come through. And I’m like, that’s totally something that, you know, once we have these muscles built.
42 00:05:18.140 ⇒ 00:05:20.210 Katherine Bayless: she could just provision, right? Like, it doesn’t…
43 00:05:20.210 ⇒ 00:05:20.630 Uttam Kumaran: Yes.
44 00:05:20.630 ⇒ 00:05:35.979 Katherine Bayless: need any further technical support if it’s using existing, you know, components that we’ve sort of defined and made available. I mean, it’s, you know, it’s like Notion CRM, but our own version of it, or, you know, tiny Salesforce, like, whatever you want to use analogy-wise, like…
45 00:05:36.040 ⇒ 00:05:43.039 Katherine Bayless: I think this is gonna be cool, if we can start to really control the nature and movement of our data.
46 00:05:43.450 ⇒ 00:05:44.989 Uttam Kumaran: Yeah. I’m like, agree.
47 00:05:44.990 ⇒ 00:05:55.160 Katherine Bayless: funny, even in the meeting earlier, like, one of the folks on Jay’s team was like, well, but, you know, we need to be able to send the data from EventPoint to Merits and back to EventPoint, and, like, how are we going to do that? And I was like, we’re just not. We never know.
48 00:05:55.160 ⇒ 00:05:55.660 Uttam Kumaran: Yeah.
49 00:05:55.660 ⇒ 00:06:08.529 Katherine Bayless: to do that, except that those systems had no central source of, like, record, and so they were just daisy-chaining stuff back and forth, and like, now we’ll just… we’ll know what the entity is and where it needs to be and when, and it will just be…
50 00:06:08.930 ⇒ 00:06:10.480 Katherine Bayless: Yeah. Yeah.
51 00:06:10.760 ⇒ 00:06:14.720 Katherine Bayless: So… Exactly. It takes some time, but, I’m very excited.
52 00:06:16.500 ⇒ 00:06:24.060 Uttam Kumaran: Amazing. Well, sorry for derailing this, Amber, but, like, it was a great con… we just had a great conversation, and, like, I think with Jay, I’m, like, kind of explaining
53 00:06:24.170 ⇒ 00:06:27.559 Uttam Kumaran: Like, the data foundation, but also, like.
54 00:06:27.710 ⇒ 00:06:35.820 Uttam Kumaran: sort of this, like, broader, like, context engineering that needs to happen. And then some people, again, are like… I think most people just…
55 00:06:35.960 ⇒ 00:06:42.669 Uttam Kumaran: aren’t gonna… don’t really need to care… they need to care more about, like, what… they need to use the AI to help them plan.
56 00:06:42.720 ⇒ 00:06:48.940 Uttam Kumaran: And be actually more open with, like, not just, like, one-shotting, like, do this, instead being like, I have this task.
57 00:06:48.990 ⇒ 00:07:04.639 Uttam Kumaran: like, think… help me think through the right skills or integrations to use, arrive at a plan, I review the plan, execute the plan, I’ll review the outputs, and then go. And instead, like, part of what I was telling Jay is, like, we’ve all been through that journey, and I think
58 00:07:04.710 ⇒ 00:07:09.690 Uttam Kumaran: part of it is, like, you couldn’t… I don’t think you could know that that’s the happy path until you…
59 00:07:09.730 ⇒ 00:07:21.880 Uttam Kumaran: you went on sort of the one-shot journey. But I also think for everybody in the org, now that they, like, kind of… part of what I was kind of… my analogy usually is, like, it would have been hard for us to explain, like, water is wet with, like, AI.
60 00:07:21.910 ⇒ 00:07:32.170 Uttam Kumaran: now that people get it, now we have to, like, show them that the right path is actually, like, moving from this, like, do this one thing, oh, it’s not, like, doing what I said.
61 00:07:32.760 ⇒ 00:07:34.709 Uttam Kumaran: Coming with the, I have this task.
62 00:07:35.090 ⇒ 00:07:38.659 Uttam Kumaran: Like, ask me questions, let me review your plan.
63 00:07:38.790 ⇒ 00:07:43.400 Uttam Kumaran: And then go ahead and execute it. And I think that’s more of, like, a training yeah.
64 00:07:43.400 ⇒ 00:07:57.030 Katherine Bayless: Yeah. Absolutely, yeah, it’s just muscle building, right? I mean, not to make it sound trivial, or like it’s gonna happen overnight, but like, yeah, it is, it’s training, it’s muscle building, it’s getting people to put their reps in, exactly what you said, we all learned this, and they can too.
65 00:07:57.030 ⇒ 00:08:03.920 Uttam Kumaran: Yes, yeah, and hopefully we can just, like, give them the answer, like, we can just give them the answer as soon as they get there, you know? It’s like…
66 00:08:04.550 ⇒ 00:08:10.239 Uttam Kumaran: Instead of having to, like, figure it out, or, like, go on Twitter, like, yeah.
67 00:08:10.240 ⇒ 00:08:33.879 Katherine Bayless: Right, right, right, right, exactly. Yeah, I mean, I definitely think, I think Jay’s already got kind of in mind, like, some, like, skills and stuff like that that would help, like, you know, to your point, like, prompting the right questions for somebody to think about, like, because I think the other piece of it is, like, seeing the components of a system that you’re gonna need, right? And being able to, like, strike that Goldilocks balance between, like, underthinking it and overthinking it.
68 00:08:33.880 ⇒ 00:08:34.650 Uttam Kumaran: Yeah.
69 00:08:34.650 ⇒ 00:08:39.430 Katherine Bayless: I have a feeling we’re gonna err on the side of the ladder, but but yeah.
70 00:08:41.780 ⇒ 00:08:42.570 Uttam Kumaran: Amazing.
71 00:08:42.740 ⇒ 00:08:45.029 Katherine Bayless: Yeah, it’s gonna be fun. It’s gonna be wild.
72 00:08:46.340 ⇒ 00:08:52.429 Uttam Kumaran: That’s great, that’s a big win to… to do. I mean, yeah, I’d, like, love to be helpful, but that’s awesome. I mean, we’ve…
73 00:08:52.780 ⇒ 00:09:03.759 Uttam Kumaran: we have done a lot of that internally here, actually. But it’s honestly being able to, like, take advantage of a process that was gonna be an RFP, and say, like, we can own it, like, it’s pretty cool.
74 00:09:04.030 ⇒ 00:09:07.299 Katherine Bayless: I know. I was like, we just saved the organization $30,000.
75 00:09:07.300 ⇒ 00:09:09.340 Uttam Kumaran: Yeah, that’s it, you know?
76 00:09:10.030 ⇒ 00:09:34.950 Katherine Bayless: Yeah, I mean, yeah, and I think, to your point, I’ll stop derailing us, but I mean, I could totally show you, too, like, I used Claude Design… well, Jay also, like, had contributed half of the designs, and then I used it for some as well, and so it was, like, Claude Design to Claude Code, and then up to, AWS Amplify, and then Aurora Serverless, because I’m a Postgres girl. But the Amplify thing is, like, that’s a cool… that’s a cool little magic trick, like, these entire, like.
77 00:09:34.950 ⇒ 00:09:40.029 Katherine Bayless: you know, web-native apps that are, like, CICD’d out of the monorepo, and somebody asked.
78 00:09:40.030 ⇒ 00:09:40.430 Uttam Kumaran: Yeah, yeah.
79 00:09:40.430 ⇒ 00:09:44.219 Katherine Bayless: And I was like, I actually don’t know, and I don’t need to. I just tell it where in the.
80 00:09:44.220 ⇒ 00:09:45.240 Uttam Kumaran: Yeah, yeah, yeah.
81 00:09:45.240 ⇒ 00:09:47.669 Katherine Bayless: is, and it… Somehow that works.
82 00:09:47.930 ⇒ 00:09:48.760 Uttam Kumaran: Yeah.
83 00:09:48.880 ⇒ 00:09:50.080 Uttam Kumaran: Exactly.
84 00:09:52.440 ⇒ 00:09:57.310 Katherine Bayless: But yes, let’s talk about getting the robot to answer the questions correctly.
85 00:09:58.820 ⇒ 00:10:05.899 Amber Lin: This is very cool. I’m starting to try and build some things in our internal platform, too, so it’s…
86 00:10:05.900 ⇒ 00:10:06.500 Katherine Bayless: Mmm. Cool.
87 00:10:06.500 ⇒ 00:10:07.120 Amber Lin: Peter.
88 00:10:07.590 ⇒ 00:10:08.370 Katherine Bayless: Yeah, yeah.
89 00:10:08.660 ⇒ 00:10:15.179 Amber Lin: Awesome. I think the main topic I wanted to talk about is…
90 00:10:15.410 ⇒ 00:10:20.380 Amber Lin: Sidebar versus semantic viewing agents.
91 00:10:20.440 ⇒ 00:10:37.449 Amber Lin: From my research, I originally thought we can query ProdMarts, just like the sidebar. Apparently, we can’t. The tool called doesn’t work. I’ve sent the question to Snowflake folks, I’m still waiting on their answer.
92 00:10:38.130 ⇒ 00:10:47.809 Amber Lin: But I think, from what I asked the robot directly, I don’t think it’s possible. So, it’s either we use semantic views for everything.
93 00:10:47.900 ⇒ 00:11:02.330 Amber Lin: Or we use a custom tool, that essentially just does text to SQL, but then we’ll have to put in guardrails, we’ll have to put in different stuff, so I don’t want to go that direction. So with semantic views.
94 00:11:02.730 ⇒ 00:11:20.199 Amber Lin: The good thing is we can add context very easily, and make sure it knows how to calculate metrics. The only thing is that it can’t join across semantic views, like it would if it were to roam freely in FrogMarts.
95 00:11:20.460 ⇒ 00:11:33.979 Amber Lin: However, I do think we don’t have that much data in prop marts. I’ve been able to add semantic views for all of them so far, so if we are able to specify
96 00:11:34.200 ⇒ 00:11:40.419 Amber Lin: which ones need to talk to each other, I’ll just add them in one place, because what… right now.
97 00:11:40.530 ⇒ 00:11:47.049 Amber Lin: for membership, I’ve added the attendance, factor in there as well, so…
98 00:11:47.240 ⇒ 00:11:50.530 Amber Lin: If Shopify needs to talk with…
99 00:11:50.860 ⇒ 00:11:53.660 Amber Lin: attendance, then we can add that in. If…
100 00:11:53.860 ⇒ 00:12:00.350 Amber Lin: Say the marketing data needs to talk with attendance, we can add that in as well, but
101 00:12:00.490 ⇒ 00:12:07.059 Amber Lin: some of them just don’t talk with each other, so it probably will be fine.
102 00:12:07.530 ⇒ 00:12:09.710 Amber Lin: So I still want to keep
103 00:12:10.140 ⇒ 00:12:27.579 Amber Lin: people using, Snowflake Intelligence, the answer quality, at least from the feedback, the few that I saw in the sidebar, the agent answered it a lot clearly and faster.
104 00:12:28.860 ⇒ 00:12:35.669 Katherine Bayless: Yeah, so I think, like, a couple sort of scattered thoughts, like…
105 00:12:36.420 ⇒ 00:12:53.109 Katherine Bayless: I definitely acknowledge, I created, the monster in the sense of, like, I mean, yeah, people really love… because I think for them, it’s like, if they can see the Streamlit app, it gives them something to have a question about. I think once they’ve gotten through that initial, like, you know, couple visits to Snowflake.
106 00:12:53.110 ⇒ 00:13:03.020 Katherine Bayless: then they probably do start to just go straight to, like, the sidebar and talk to it, but, like, the initial sort of lightbulb moment tends to come from, like, seeing something and then being like, oh, I have a question.
107 00:13:03.100 ⇒ 00:13:19.739 Katherine Bayless: I mean, I think, you know, to a certain extent, it’s worth, like, you know, reminding for the good of the order, right? Like, folks here have been data-starved for so long that, like, I mean, they think it’s impressive when I can tell them the top 10 countries at CES, right? Like, they…
108 00:13:19.740 ⇒ 00:13:20.200 Amber Lin: Mmm.
109 00:13:20.540 ⇒ 00:13:25.880 Katherine Bayless: A lack of imagination that is a result of how hard it has been to go.
110 00:13:25.880 ⇒ 00:13:26.550 Amber Lin: Hold the data.
111 00:13:26.970 ⇒ 00:13:43.230 Katherine Bayless: And so we’re really trying, like, I’m just trying anything and everything to get them to, like, engage and also understand the power. And so, yeah, so the, like, see something, ask something had been really sticky, and so I’ve created the problem of them really liking those two things together.
112 00:13:43.230 ⇒ 00:13:48.950 Amber Lin: I completely understand. I think that’s important for user adoption. It’s like, something’s better than
113 00:13:49.060 ⇒ 00:14:00.390 Amber Lin: than not using it. I did research that, and we should be able to call the agent API, through Streama apps, so maybe that’s more…
114 00:14:00.390 ⇒ 00:14:00.910 Katherine Bayless: Hmm.
115 00:14:00.910 ⇒ 00:14:17.920 Amber Lin: direct way to say, hey, you can just ask, hey, if you have questions about this dashboard, you can ask the agent something. The interface, it will probably be slower than directly creating snow intelligence, but we can experiment with that, or at least
116 00:14:18.330 ⇒ 00:14:25.279 Amber Lin: Put a link to Snow Intelligence in the Streamla apps, if that’s where they’re usually at.
117 00:14:25.430 ⇒ 00:14:28.540 Amber Lin: And if they get used to it, they’ll just go there.
118 00:14:28.540 ⇒ 00:14:37.069 Uttam Kumaran: Yeah, I think my… one of my pieces is, one is, like, I think part of this is, like, just that we’re learning this brand new technology, so I think…
119 00:14:37.170 ⇒ 00:14:43.649 Uttam Kumaran: we… I think, Amber, we almost need to have one system that, like, even if it’s…
120 00:14:43.680 ⇒ 00:14:58.340 Uttam Kumaran: the worst long-term solution is just, like, all reliable, and I think if that is just a sidebar, then we just, like, try to make sure that that, like, works. I do think that Snowflake intelligence in, like, a layer that is more like co-work.
121 00:14:58.340 ⇒ 00:15:07.339 Uttam Kumaran: within Snowflake is probably the future, and I think we kind of have to hold both of those true, but what I don’t want us to do is, like.
122 00:15:07.440 ⇒ 00:15:20.809 Uttam Kumaran: get distracted and, like, develop the next new surface, and then kind of, like, not finish the tackle on something that we can actually use and continue to gain, sort of, support with now.
123 00:15:20.950 ⇒ 00:15:29.849 Uttam Kumaran: So, I don’t know if that sort of kind of summarizes a little bit, because that’s my feeling as well. It’s like, I do think that it feels like intelligence will be the, like.
124 00:15:30.220 ⇒ 00:15:40.839 Uttam Kumaran: basically the… the… they will continue to invest in probably that product, I’m thinking. And then also our ability to build custom on top is probably even, like, the next frontier.
125 00:15:40.840 ⇒ 00:15:53.219 Uttam Kumaran: It’s like bringing this into other applications, hitting Snowflake directly, but I also agree, like, if people are used to Sidebar now, and it’s sort of something we just want to nail, then we should knock that out, and then we should basically
126 00:15:53.250 ⇒ 00:15:59.859 Uttam Kumaran: I think, Catherine, the biggest thing is, like, if people hit… there are some technical limitations that we’re finding, how can we…
127 00:15:59.970 ⇒ 00:16:06.189 Uttam Kumaran: Basically use those to pivot people into a new surface, if we want to do that.
128 00:16:08.430 ⇒ 00:16:10.470 Katherine Bayless: Yeah, I think…
129 00:16:11.040 ⇒ 00:16:14.979 Uttam Kumaran: You can also say, like, okay, I don’t think that’s right, like, I don’t… you know, I’m just…
130 00:16:15.250 ⇒ 00:16:15.860 Katherine Bayless: We’re sticking through.
131 00:16:15.860 ⇒ 00:16:16.830 Uttam Kumaran: True, yeah.
132 00:16:17.090 ⇒ 00:16:38.499 Katherine Bayless: I mean, I think… no, I very much agree, because I was going to say, like, I think the other problem that I have caused, which is funny, because I’ve also, like, been so belligerently soapboxing about this around the organization, is, like, we don’t have a staging environment, right? Like, I keep telling everybody else they should have one, and then here I am not walking my own talk. And so, like, I think that’s been…
133 00:16:38.500 ⇒ 00:16:41.620 Katherine Bayless: In retrospect, one of the other things is, like, we’re trying to, like.
134 00:16:41.660 ⇒ 00:16:56.429 Katherine Bayless: do the semantic models in production, where people are already working, and then it’s like, the experience they had yesterday isn’t the one they’re having today. And it’s not, like, any fault or signal that the semantic model is not wrong, it’s just the problem is that we’re deploying in production, right? Like, I need a.
135 00:16:56.430 ⇒ 00:16:56.990 Uttam Kumaran: Yeah.
136 00:16:56.990 ⇒ 00:17:16.900 Katherine Bayless: I need to have us actually have a place where we can do safe staging, so that we can do all the refining with a QA team before, you know, we finally ship something out. So that’s, I mean, yeah, to be fair, totally another one on me. Yeah, like, we don’t have a safe place to learn that it’s gonna say, I don’t know about CES before I show it to Gary’s assistants.
137 00:17:17.670 ⇒ 00:17:21.779 Katherine Bayless: I think, too, like…
138 00:17:21.890 ⇒ 00:17:36.429 Katherine Bayless: it’s interesting, the semantic views and the joins thing, like, joining across them, I’m inclined to think, Utam, you’re probably right, that, like, Snowflake’s gotta have a plan for this, because another sort of, like…
139 00:17:36.570 ⇒ 00:17:51.280 Katherine Bayless: trick we don’t want to fall into is, like, part of what I’m campaigning, you know, like, all your data in here, and then we can have it all in one place, and we can ask questions across it, and… I mean, they’re probably going to ask questions that go across data sources that we would not.
140 00:17:51.280 ⇒ 00:17:52.040 Uttam Kumaran: Totally.
141 00:17:52.400 ⇒ 00:18:05.890 Uttam Kumaran: Some of that is, like, if Snowflake doesn’t allow for it, then we will just build a custom tool. And so I want to leave that on the table, because I feel we’re more than comfortable to do that. So if we feel, Amber, like we’ve hit
142 00:18:05.890 ⇒ 00:18:19.820 Uttam Kumaran: sort of a limit, or we don’t have a path forward, we should just… we could just ship our own little MCP or tool that… that does this, because straight text to SQL with a semantic layer is pretty… it’s pretty good, and we can do that. Like, I don’t think it’s…
143 00:18:19.820 ⇒ 00:18:31.519 Uttam Kumaran: it’s not, like, some giant project for us to do, if it allows for some functionality. I would love for SoFi to come back to us and sort of enable that, but I also think that
144 00:18:31.570 ⇒ 00:18:39.029 Uttam Kumaran: that’s something probably necessary we would have to do, if we look to integrate Snowflake and querying Snowflake into other
145 00:18:39.030 ⇒ 00:18:52.180 Uttam Kumaran: things outside of Snowflake itself, right? Because right now, we’re all building for what I’m calling the surface that is, like, Snowflake intelligence, not… not, like, something that’s external. So, maybe it’s worth us exploring that,
146 00:18:52.870 ⇒ 00:19:06.480 Uttam Kumaran: Yeah, but you’re… but you’re totally right on the staging layer, so I think, Amber, that’s also something maybe you can take on, is, like, can… how do we create, like, an area where we can have these semantic views go to staging? I think Awash can totally help you solve that.
147 00:19:08.960 ⇒ 00:19:22.409 Amber Lin: I was developing in dev before, but I think with the sidebar, the downside is as long as everything is… it can… it can see everything, so never… no matter what folder I am in, it can see.
148 00:19:22.410 ⇒ 00:19:23.010 Katherine Bayless: Right.
149 00:19:23.010 ⇒ 00:19:25.170 Amber Lin: every single thing, so it’s just niece.
150 00:19:25.170 ⇒ 00:19:25.730 Uttam Kumaran: Yeah.
151 00:19:25.730 ⇒ 00:19:27.130 Amber Lin: different. Maybe that.
152 00:19:27.130 ⇒ 00:19:27.520 Uttam Kumaran: I mean.
153 00:19:27.520 ⇒ 00:19:27.980 Amber Lin: with Sam.
154 00:19:27.980 ⇒ 00:19:28.320 Uttam Kumaran: Nope.
155 00:19:28.320 ⇒ 00:19:30.059 Amber Lin: sandbox was for, but we don’t.
156 00:19:30.060 ⇒ 00:19:35.559 Uttam Kumaran: Yeah, we could replicate it all on Sandbox, and not delete it.
157 00:19:35.560 ⇒ 00:19:50.230 Katherine Bayless: It’s true, so it’s funny, the sandbox, it was only set up because, I feel, you know, I’m not picking on him. When I first, bought Snowflake, Jay, I asked him to set up the SSO for it, and he went down the path of doing it, and, like, ran into a funny, like.
158 00:19:50.230 ⇒ 00:19:55.990 Katherine Bayless: I don’t know, there was this weird, like, chicken or egg loop, and he hit it wrong, and we basically got bricked out of Snowflake, and so in the.
159 00:19:55.990 ⇒ 00:19:56.610 Uttam Kumaran: Okay, yeah.
160 00:19:56.610 ⇒ 00:20:18.579 Katherine Bayless: like, I need to figure out what that was supposed to work like, and so he set up the sandbox, figured out how it was supposed to work, and then we got unbricked on the prod one. But, to your point, Ambert, we could totally actually have a staging instance, because I think another one that’s on me is, like, yes, the sidebar thing can see everything, and when I’m demoing, I’m still in role developer, like, I’m not necessarily, like.
161 00:20:18.580 ⇒ 00:20:19.090 Uttam Kumaran: Yeah.
162 00:20:19.090 ⇒ 00:20:36.470 Katherine Bayless: remembering to toggle over to role chat, because usually I’m trying to show them, like, here’s something I built for you that’s not quite, you know, in production yet, but, like, I want to demo it to you, but then I’m in role developer, and the sidebar chat sees things that are not, like, you know, what we want it to be seeing in that demo context, and so, yeah, it’s like, I’m just…
163 00:20:36.470 ⇒ 00:20:53.700 Katherine Bayless: I’ve mixed everything in a place and made it hard in that way, and so I think we probably still want to close that sandbox instance, because it’s tied to the old AWS accounts anyway. Okay. But we could totally, set up a real staging sandbox snowflake, and then also still have our production one.
164 00:20:53.890 ⇒ 00:21:12.670 Uttam Kumaran: Yeah, Amber, can you ask, thanks to that as well, in our chat? Like, see what he says about that, because, yeah, I mean, that seems like the clearest alternative. I mean, Katherine, we’re seeing this across a few clients where the demand for the solution is so high that, like, the urge to demo, like.
165 00:21:13.330 ⇒ 00:21:16.460 Uttam Kumaran: actually causes… then it’s like, oh shit, we’re live now.
166 00:21:16.690 ⇒ 00:21:18.049 Katherine Bayless: Right, right.
167 00:21:18.050 ⇒ 00:21:23.889 Uttam Kumaran: And I also want to just make sure that, like, we say that out loud, and that we’re, like.
168 00:21:24.160 ⇒ 00:21:35.379 Uttam Kumaran: careful, or, like, we… we just, like, treat, like… for example, I think on our side, maybe… I think we should have treated your meeting with them with, like, okay, let’s get Catherine, like, not a, like, a…
169 00:21:35.860 ⇒ 00:21:44.510 Uttam Kumaran: like, a hard-coded script, but, like, here’s things we’re, like, pretty confident in, and take those moments like…
170 00:21:44.640 ⇒ 00:21:49.989 Uttam Kumaran: You know, to just nail that, because we’re also now, like, we went live… we kind of live with the product, you know?
171 00:21:49.990 ⇒ 00:21:50.800 Katherine Bayless: Oh yeah, we’ve been liking.
172 00:21:50.800 ⇒ 00:22:00.010 Uttam Kumaran: We didn’t go through, like… Yeah, yeah, we didn’t go, like, we didn’t do, like, a six-month UAT, like, you know, so part of this is gonna be… there will be a couple hiccups.
173 00:22:00.270 ⇒ 00:22:08.180 Uttam Kumaran: But also, like, the demand is so high, and everything is happening all at once, so we need to still try to have some type of software lifecycle.
174 00:22:09.300 ⇒ 00:22:20.030 Katherine Bayless: Yeah, I mean, yeah, agreed, agreed. Like, I think the demand was high, and then the interaction and engagement was swift, right? I mean, it was like, you know, like.
175 00:22:20.030 ⇒ 00:22:20.720 Uttam Kumaran: Yes.
176 00:22:20.720 ⇒ 00:22:34.860 Katherine Bayless: spread like wildfire. I mean, I went from, like, you know, inviting in the handful of folks from membership that we were working with, and then, like, you know, two people that I was like, yeah, you guys will be early adopters. And, you know, now we’ve got over, like, 50 people in there, and they’re loving it, right? Like, and so I’m…
177 00:22:34.860 ⇒ 00:22:47.499 Katherine Bayless: I’m not mad at that, because I am totally not the type of person who tends to scope big, you know, rollouts and sort of those sorts of things, but to your point, we’re starting to kind of graduate from the, you know, chaos stage. Like, if we’re gonna be…
178 00:22:47.500 ⇒ 00:22:58.339 Katherine Bayless: managing disposable software, too, like, yeah, a little bit more of a buttoned-up posture is appropriate, instead of the three toddlers in a trench coat that we’ve been behaving like. But.
179 00:22:58.340 ⇒ 00:23:06.949 Uttam Kumaran: Yeah, so that’s why I wonder, like, how much of that you think we can buy? Like, how much, you know,
180 00:23:08.360 ⇒ 00:23:12.450 Uttam Kumaran: How much of, like, space… He said we’re gonna…
181 00:23:13.240 ⇒ 00:23:19.680 Uttam Kumaran: On this, like, do you think there are milestones we can try to set, and that way it gives us some time in between?
182 00:23:21.230 ⇒ 00:23:23.400 Katherine Bayless: I mean, I think…
183 00:23:28.190 ⇒ 00:23:29.679 Uttam Kumaran: I mean, it’s hard, it’s a tough question.
184 00:23:29.680 ⇒ 00:23:36.439 Katherine Bayless: No, it’s an interesting question, I mean, because it kind of pokes at, like, the philosophical core here is, like.
185 00:23:36.680 ⇒ 00:23:49.830 Katherine Bayless: I mean, okay, so taking a step back, when I interviewed, right, like, the people that were seated on the other side of the table, you know, they were like, data governance, center of excellence, roadmap, right? Like, they wanted the plan, they wanted…
186 00:23:49.830 ⇒ 00:23:50.409 Uttam Kumaran: Best, yeah.
187 00:23:50.410 ⇒ 00:23:52.580 Katherine Bayless: governance to be the leading thing, and I would.
188 00:23:52.580 ⇒ 00:23:53.400 Uttam Kumaran: Yeah.
189 00:23:53.400 ⇒ 00:23:55.569 Katherine Bayless: it’s a weaponized stall tactic. Like, I’m.
190 00:23:55.570 ⇒ 00:23:56.360 Uttam Kumaran: Yeah, yeah, yeah.
191 00:23:56.360 ⇒ 00:24:00.690 Katherine Bayless: wildfire, and I would just get people engaging with stuff, and so, like.
192 00:24:01.110 ⇒ 00:24:13.409 Katherine Bayless: I think that is probably likely to continue to be the, like, modality of, like, intake, right? Like, I think that’s where I’m… I’m starting to see we can flip the model and, like, get people in and engaging.
193 00:24:13.410 ⇒ 00:24:24.889 Katherine Bayless: but understanding, like, certain things are, like, polished, and certain things are under construction, and certain things are not yet ready, right? Like, like, I think it’s gonna be…
194 00:24:25.670 ⇒ 00:24:36.419 Katherine Bayless: more beneficial for the way this organization seems to move, like, to have them understand, like, the gradient of readiness of the data, rather than a
195 00:24:36.420 ⇒ 00:24:47.969 Katherine Bayless: process with milestones and, like, you know, kickoffs and deliverables, like, I think there will be instances where, like, kickoffs and milestones make sense, but for the most part, I think this is gonna be…
196 00:24:47.990 ⇒ 00:25:04.579 Katherine Bayless: kind of continue to be a building the plane as we fly it, and educating folks to understand how to tell the difference between, like… and I mean, not to say that it’s, like, gonna be a mystery, but, like, how to engage with our data, knowing that some of it has been thoroughly vetted, some of it’s in flight, some of it’s, you know… it’s in there!
197 00:25:04.580 ⇒ 00:25:13.310 Katherine Bayless: but not really ready. So yeah, like, I think that’s the muscle building I want to move towards, rather than the cadence, you know?
198 00:25:13.930 ⇒ 00:25:27.629 Uttam Kumaran: Yeah, I just think, like, for, like, one kind of anecdote on our side, like, we had a startup we were helping with, and we shipped for them, like, the first version of, like, AI data analysis within their BI tool, but then it got so much adoption because they had, like, no reporting.
199 00:25:27.630 ⇒ 00:25:41.459 Uttam Kumaran: But then we just got flooded with, like, this is wrong, this is wrong, this is wrong, and there was no space to breathe, and this is wrong isn’t necessarily, like, a one-line fix. It can be, like, something systemic.
200 00:25:41.550 ⇒ 00:25:49.060 Uttam Kumaran: You know, something kind of deep, and so… and then those questions, they’re… now the ease of asking a query is now…
201 00:25:49.240 ⇒ 00:26:01.470 Uttam Kumaran: whenever you feel like opening stuff and asking a query, that, like, if they ask 5 days in a row and it doesn’t work, do we lose that person? Now, I think CTA is a different environment than, sort of, like, a startup, but…
202 00:26:01.600 ⇒ 00:26:05.130 Uttam Kumaran: I just wanna… seeing that happen just now.
203 00:26:05.400 ⇒ 00:26:18.700 Uttam Kumaran: reminded me to, like, hey, we actually should treat this like a product rollout. I’m kind of with you in that I’m also not, like, a huge 100 milestones, like, you know, type of person, but,
204 00:26:19.420 ⇒ 00:26:25.589 Uttam Kumaran: I would say, like, for the benefit of everybody, trying to think through some release cadence.
205 00:26:25.840 ⇒ 00:26:29.750 Uttam Kumaran: Where, and then, like, kind of communicating expectations to people.
206 00:26:29.900 ⇒ 00:26:42.840 Uttam Kumaran: is gonna be helpful, whether that’s every week, whether that’s every two weeks, whether, you know, we were… but, like, again, you can push back and say, like, everybody here gives us a lot of benefit to the doubt, they’ll keep trying until it works. Because that’s kind of our company, like.
207 00:26:42.860 ⇒ 00:26:54.939 Uttam Kumaran: they know we’re trying to build things, that everybody has a lot of benefits and doubt, and people are perseverant, they’re not like, this doesn’t work. But again, if you’re in a heavy politics environment, and that could degrade trust, like.
208 00:26:55.110 ⇒ 00:27:02.519 Uttam Kumaran: that’s what I want to make sure, because some of… for example, like, moving out of semantic views to this, it’s not a 24-hour turnaround.
209 00:27:02.690 ⇒ 00:27:08.960 Uttam Kumaran: It’s like… it’s just gonna take some time, and so that’s what I wanna just, like, think about.
210 00:27:09.800 ⇒ 00:27:14.929 Katherine Bayless: Yeah, I mean, I think… so, like, another example, though, is the disposable software thing, like, the…
211 00:27:14.930 ⇒ 00:27:15.690 Uttam Kumaran: Yes.
212 00:27:15.690 ⇒ 00:27:34.519 Katherine Bayless: there’s this AI task force proposal document that I’ve done three drafts of, and it’s still under review somewhere. Meanwhile, we’ve already greenlit a full-on external-facing use case, right? And so, like, our ability to plan is lagging behind our ability to execute, and I don’t want to get bogged down in the planning, but I do hear.
213 00:27:34.520 ⇒ 00:27:34.840 Uttam Kumaran: Yeah.
214 00:27:35.060 ⇒ 00:27:36.690 Katherine Bayless: And I think…
215 00:27:36.790 ⇒ 00:27:49.869 Katherine Bayless: we also… we also know that we need to start scaling this sort of, like, interaction across the team, so that it’s not only the data team that can, like, you know, opine and provide the yes-no’s. Yeah.
216 00:27:49.870 ⇒ 00:28:00.529 Katherine Bayless: I think we also know that we need to start onboarding, you know, a number of new data sources as the year starts to progress, but we’ve kind of tackled the, like, really gnarly ones that are very central.
217 00:28:00.530 ⇒ 00:28:20.989 Katherine Bayless: I think people in more, sort of, like, senior roles are going to be the ones that are in that space you’re talking about, Utam, that are like, if they come in, they get a couple wrong answers, we’re gonna lose them, right? Because all they’re gonna do is go in to tell their team to do it instead of the robot, and so, like, they, I think, are a more fragile audience.
218 00:28:20.990 ⇒ 00:28:26.050 Katherine Bayless: the people that are on the ground, they absolutely… like, right now, I mean, their jobs are, like.
219 00:28:26.050 ⇒ 00:28:27.210 Uttam Kumaran: They’ll sign whatever.
220 00:28:27.210 ⇒ 00:28:30.400 Katherine Bayless: All day, every day, and so, like, if we can have them
221 00:28:30.400 ⇒ 00:28:55.389 Katherine Bayless: you know, still use that sort of, you know, time and brainpower, but, like, QAing a semantic view, right? And then explaining, like, once this is done, all of the reporting you’re doing, you know, we will automate it through these pipelines. And so, like, I think we can lean on the, like, on-the-ground folks to do a lot more of the review and refining and QA, and then, yeah, to your point, maybe with these sort of more senior folks.
222 00:28:55.390 ⇒ 00:28:56.240 Katherine Bayless: are…
223 00:28:56.920 ⇒ 00:29:12.309 Katherine Bayless: doing some sort of release cadence in the sense of, like, announcing, you know, this data source is now ready for, you know, ready and polished, and, like, this one is still in QA, but we expect it to deliver in a couple weeks. Like, I think that kind of information does make sense to push out.
224 00:29:12.360 ⇒ 00:29:18.580 Katherine Bayless: But I wouldn’t want them to come back and, like, you know, be like, okay, well, what’s your, you know, what’s your 5-year plan? Like, I plan 5 minutes.
225 00:29:18.580 ⇒ 00:29:19.360 Uttam Kumaran: Yes.
226 00:29:20.030 ⇒ 00:29:21.809 Uttam Kumaran: Yeah, yeah, okay, okay.
227 00:29:25.640 ⇒ 00:29:29.830 Katherine Bayless: There was something else I was gonna say. I forget what it was, though. But yeah, Amber, go ahead, you came off mute.
228 00:29:30.420 ⇒ 00:29:37.890 Amber Lin: No, I was… I’m interested in this conversation, I just…
229 00:29:39.300 ⇒ 00:29:57.690 Amber Lin: I will go with whatever we decide on, in terms of rollout, and happy to call people to support them in terms of, hey, this is what the semantic view is like, so if you have power users, you can connect me with them, and I can go with them one-on-ones on these things.
230 00:29:57.960 ⇒ 00:30:02.270 Amber Lin: And… Let’s see… Yeah.
231 00:30:02.270 ⇒ 00:30:09.549 Uttam Kumaran: Catherine, is there… is there anyone that comes to mind that we can start to have Amber maybe take some of your load off that is doing those
232 00:30:09.710 ⇒ 00:30:15.629 Uttam Kumaran: like, sort of one-on-one sessions, and also it’ll give Amber sort of direct feedback, too.
233 00:30:16.860 ⇒ 00:30:39.459 Katherine Bayless: Yeah, so there… there are a few people, in fact, so, like, the membership stuff is a good example. So, like, Anna Rutter’s very much been sort of, like, our, you know, early champion, and kind of, like, leading her little team, engaging with us on the building out of all the Streamlit apps and stuff like that. But then, in addition to her, there are a couple people on her team who have a lot of the heavy, like, reporting,
234 00:30:39.460 ⇒ 00:30:43.890 Katherine Bayless: Responsibilities, and so, like, working with them, I think would make sense.
235 00:30:43.890 ⇒ 00:30:54.379 Katherine Bayless: I do want to do a, like, a, like, maybe can we have a technical deep dive moment on the whole semantic views joining to other semantic views thing? Because I do…
236 00:30:54.400 ⇒ 00:31:09.850 Katherine Bayless: I do wonder, like, that AMS data share, because it’s an entire CRM, I assume it needs more than one semantic view, but then if those can’t talk to each other, like, help me understand how the different semantic views can still.
237 00:31:09.850 ⇒ 00:31:10.390 Uttam Kumaran: Yeah.
238 00:31:10.390 ⇒ 00:31:11.240 Katherine Bayless: together.
239 00:31:12.670 ⇒ 00:31:30.339 Amber Lin: So in the semantic view, in Snowflake, they recommend that we keep it under 20 tables. All of our semantic views are under 20 tables, because our… it’s usually, like, 11 or so. The current CRM is about 9 tables, I believe, and…
240 00:31:30.590 ⇒ 00:31:39.670 Amber Lin: When they query, because we’ve defined the relationships between table, they can join together. Outside of the semantic view.
241 00:31:39.900 ⇒ 00:31:59.680 Amber Lin: the agent wouldn’t know, hey, this table joins to this and this way, and so it wouldn’t do that, and I don’t think it has permission to. The only thing would be they run separate… two separate queries, and then put them together, or maybe take the result of one query, and then say, hey, this is the
242 00:32:00.360 ⇒ 00:32:11.429 Amber Lin: create this query based on… based on this. If we wanted to talk to each other, one would be creating a bigger semantic view that covers all of them.
243 00:32:11.850 ⇒ 00:32:13.250 Katherine Bayless: Yeah.
244 00:32:13.250 ⇒ 00:32:19.309 Amber Lin: You can create a model that’s just through engineering, create a model that talks to each other.
245 00:32:20.680 ⇒ 00:32:28.679 Amber Lin: And… but I think if they’re connected, they should be in one semantic view, because we need that to answer questions.
246 00:32:28.680 ⇒ 00:32:32.530 Uttam Kumaran: Yeah, so I wonder if the first… I mean, my initial take is just, like, yeah, I…
247 00:32:32.740 ⇒ 00:32:36.840 Uttam Kumaran: 20 tables is not, like, a lot. I think,
248 00:32:37.050 ⇒ 00:32:47.720 Uttam Kumaran: I’m surprised to hear, like, that’s the recommendation, but then they don’t allow for, like, querying multiple semantic views. Like, if that’s the case, then I would just say, Amber, we do need to just…
249 00:32:48.480 ⇒ 00:32:50.149 Uttam Kumaran: throw that.
250 00:32:51.280 ⇒ 00:33:07.470 Katherine Bayless: Yeah, because, like, we’ve only scratched the surface of the stuff, like, we pulled out the, like, the obvious, like, we need these now pieces from that CRM, but there’s so much more in there we need to bring in, and I think, yeah, I mean, it could absolutely be more than 20 tables, but it could be more than 20 tables, but…
251 00:33:07.710 ⇒ 00:33:09.810 Katherine Bayless: Slightly different, like…
252 00:33:10.320 ⇒ 00:33:15.570 Katherine Bayless: I mean, it’s all membership data, but I’m like, I guess you could break it up in, like.
253 00:33:16.120 ⇒ 00:33:20.509 Katherine Bayless: memberships versus content? No, I don’t know, I don’t know, actually, yeah, I mean…
254 00:33:20.900 ⇒ 00:33:24.929 Uttam Kumaran: Well, I also think, like, yeah, this is where it’s maybe an awaits question on, like.
255 00:33:25.040 ⇒ 00:33:36.230 Uttam Kumaran: can you keep your marts? Well, this is also where, I think, Catherine, this more goes more to, like, a number of tables is sort of not… like, I don’t think in terms of number of tables, it’s like…
256 00:33:36.340 ⇒ 00:33:49.019 Uttam Kumaran: okay, should we just have wider tables and the number goes down? Like, you know, what is the… what is actually, like, the limiting factor here? So ultimately, like, this is where I think it goes back to Amber, like, if…
257 00:33:49.020 ⇒ 00:34:01.779 Uttam Kumaran: if the recommendation is, like, to keep under a certain amount of tables, I want to know from Snowflake, like, what are they actually trying to say here? Like, because that doesn’t mean anything necessarily to me. Like, are they saying, like, okay, after, like.
258 00:34:01.880 ⇒ 00:34:13.620 Uttam Kumaran: 600 columns, like, there’s some thing, after a certain, like, side limit, we’re seeing degradation. Because ultimately, like, we can combine those tables into fatter tables, and then you hit your limit. So this is also where I think
259 00:34:13.800 ⇒ 00:34:19.149 Uttam Kumaran: what we should do, Amber, shorter term, is, like, we should just have bigger semantic views.
260 00:34:19.340 ⇒ 00:34:20.320 Uttam Kumaran: I would like…
261 00:34:20.639 ⇒ 00:34:29.760 Uttam Kumaran: to have some unders… like, and then I want to… I can also double down on your message to Snowflake, basically trying to get an understanding of, like, what they mean by 20 tables.
262 00:34:30.179 ⇒ 00:34:33.010 Uttam Kumaran: And then second is, like, my hunch…
263 00:34:33.179 ⇒ 00:34:39.010 Uttam Kumaran: Is that we need to move to a model that’s wider tables, and actually, like.
264 00:34:39.429 ⇒ 00:34:45.300 Uttam Kumaran: Catherine, is what our conversation, like, the notion of scar schema was… fine when…
265 00:34:45.449 ⇒ 00:34:56.159 Uttam Kumaran: you were… humans were sort of querying and… and, you know, curating for organization. Now that’s, like, actually a lot less relevant.
266 00:34:56.400 ⇒ 00:35:02.679 Uttam Kumaran: Like… And maybe the thing that’s still relevant is just, like, yeah, fatter tables.
267 00:35:02.810 ⇒ 00:35:17.649 Uttam Kumaran: maybe perform less well when queried, but, like, you can handle that through views and, like, a lot of different ways, like, I kind of… that’s, like, my… if I was to, like, put my money on something, it’s probably what we need to do, and, like.
268 00:35:17.990 ⇒ 00:35:23.240 Uttam Kumaran: I don’t think I… we… yeah, this is, like, so insane, I just never thought… I just didn’t think that, like, that…
269 00:35:23.570 ⇒ 00:35:30.099 Uttam Kumaran: Yeah, like, but that makes a lot more sense, because the AI just pulls the entire DDL into context and runs whatever query it needs.
270 00:35:30.960 ⇒ 00:35:31.370 Katherine Bayless: you know.
271 00:35:31.370 ⇒ 00:35:33.660 Uttam Kumaran: It’s that extra hop to go figure out
272 00:35:33.940 ⇒ 00:35:37.609 Uttam Kumaran: The relevant table, blah blah blah, like, we should just put it all into one.
273 00:35:38.260 ⇒ 00:35:48.859 Katherine Bayless: Yeah, and I mean, I think the other thing that works to our advantage, right, is that, like, I mean, at the end of the day, at least at present, who knows, maybe someday it’ll be different, but, like, we have tiny data, right? I mean, we… tiny, tiny data.
274 00:35:48.860 ⇒ 00:35:49.430 Uttam Kumaran: Yeah, yeah, yeah.
275 00:35:49.430 ⇒ 00:35:55.329 Katherine Bayless: You know, we’re not dealing with, like, petabytes of stuff where, like, a wide table is a death sentence for the compute, right?
276 00:35:55.330 ⇒ 00:35:55.720 Uttam Kumaran: Yes.
277 00:35:55.720 ⇒ 00:36:00.919 Katherine Bayless: you know, whether we go wide or star schema, like, I was just trying to be a good dev and suggest star schema.
278 00:36:00.920 ⇒ 00:36:01.500 Uttam Kumaran: Yeah.
279 00:36:01.500 ⇒ 00:36:04.729 Katherine Bayless: But now I’m like, I think that’s not how this works anymore.
280 00:36:04.730 ⇒ 00:36:05.430 Uttam Kumaran: Yeah.
281 00:36:05.430 ⇒ 00:36:11.390 Katherine Bayless: enough to go wide and fat with it. Like, I don’t know, I really think, like.
282 00:36:11.400 ⇒ 00:36:15.640 Katherine Bayless: And actually, there’s another piece you were talking about with the snowflake stuff,
283 00:36:15.650 ⇒ 00:36:24.159 Katherine Bayless: I actually think it might be worth exploring… Amber, you had mentioned this, I think, in the Slack thread or somewhere else, but, like, maybe we start with just…
284 00:36:24.160 ⇒ 00:36:37.200 Katherine Bayless: putting the table metadata in place, so, like, defining the fields. I had done an early, you know, sort of, you know, faux semantic view on the CES Reg data back when, Dave Hennessy started asking questions, and I, like.
285 00:36:37.200 ⇒ 00:36:51.869 Katherine Bayless: defined the, like, 3 or 4 fields on the table that were most relevant, and then in the, like, you know, table summary at the top, I was like, in order to tell a user what a verified attendee is, you know, this is the logic, and then I gave it a few other, like, things, and then I was like.
286 00:36:51.870 ⇒ 00:37:09.149 Katherine Bayless: performance increased dramatically, and I was like, good enough, moving on. We’ll do semantic views later. But I’m like, I wonder if we go back to working on table metadata and trying to put those definitions in that way, and see if maybe Snowflake’s gonna solve this problem for us, because I just can’t imagine
287 00:37:09.150 ⇒ 00:37:26.150 Katherine Bayless: we’re the only people who are sitting here going, like, I think the robot’s better at our data than we are at explaining it, right? And so I don’t know, I don’t know, that’s an idea that I had around, like, maybe it’s a way to scale back and stop feeling like we’re doing the right thing but getting a weird result.
288 00:37:26.150 ⇒ 00:37:31.209 Katherine Bayless: And also, you know, I don’t know. We’d learn along the way, and it needs to happen anyway.
289 00:37:34.140 ⇒ 00:37:34.619 Amber Lin: Yeah, I agree.
290 00:37:34.620 ⇒ 00:37:36.550 Uttam Kumaran: Yeah, so I… You should do that.
291 00:37:37.170 ⇒ 00:37:41.410 Amber Lin: I… so that’s… I think that’s the only way we can…
292 00:37:42.170 ⇒ 00:37:45.000 Amber Lin: Be in the sidebar and trust
293 00:37:45.130 ⇒ 00:37:53.829 Amber Lin: When it’s not using semantic views, because in the sidebar, it can’t control unless it explicitly copied the name of the semantic view.
294 00:37:54.690 ⇒ 00:37:57.550 Amber Lin: It will wander off, and…
295 00:37:57.710 ⇒ 00:38:03.750 Amber Lin: In those cases, I have no way to say, hey, this column means this.
296 00:38:04.630 ⇒ 00:38:19.580 Katherine Bayless: Yeah, I mean, the funny thing is, like, it tends to wander off pretty brilliantly, like, I one-shotted a bank rec ledger at one point, like, I just told it, I was like, hey, out of the remembers data, can we make a bank reconciliation report for finance? They’d love that, and it…
297 00:38:19.620 ⇒ 00:38:29.430 Katherine Bayless: It thought for a while. But, like, it figured it all out, right? And, like, I mean, I’ve also asked it questions where it’s come back with, like, you know, having brought back data from a different place, and I’m like, oh, I didn’t…
298 00:38:29.580 ⇒ 00:38:49.409 Katherine Bayless: I didn’t really think to ask that, or, like, I didn’t think that you’d find that there, but that’s interesting. Like, I don’t know, I don’t want to compromise on the, like, serendipity of some of the connections it finds, especially as we do start bringing a bunch of, sort of, like, disparate stuff that’s never really been connected before in. And maybe if we define all the table metadata and…
299 00:38:49.410 ⇒ 00:38:59.199 Katherine Bayless: That makes it a little easier to pull a semantic view together, because now you’re able to just kind of, like, harvest it right off of the table, versus needing to have it come through, like, a separate channel.
300 00:38:59.200 ⇒ 00:38:59.810 Uttam Kumaran: Yeah.
301 00:39:01.630 ⇒ 00:39:11.240 Uttam Kumaran: Yeah, so maybe, Amber, you want… maybe you could do that, like, I think there’s two changes here. I think you should prioritize backfilling, like, metadata wherever we have it. I think
302 00:39:11.910 ⇒ 00:39:18.860 Uttam Kumaran: I think, Catherine, the challenge here is, like, some of that we may not have the definitions of, or, like, so…
303 00:39:19.470 ⇒ 00:39:20.670 Uttam Kumaran: I think we have to…
304 00:39:21.100 ⇒ 00:39:30.139 Uttam Kumaran: Like, how we want to split, sort of, the responsibility there, because we can do it from what we have, and, like, we have some transcripts and documents, but…
305 00:39:30.340 ⇒ 00:39:35.859 Uttam Kumaran: We may just have to do a powwow on, like, writing all of it down in Excel so that Amber can backfill it.
306 00:39:37.330 ⇒ 00:39:39.379 Katherine Bayless: Yeah. I think.
307 00:39:39.380 ⇒ 00:39:41.219 Uttam Kumaran: Otherwise, it’ll hallucinate it, you know?
308 00:39:41.220 ⇒ 00:39:47.770 Katherine Bayless: Right, yeah, and I was gonna say, like, I know Kyle has a few times been like, look, it wrote it for me, and I’m like, no, we’re not gonna… this is actually.
309 00:39:47.770 ⇒ 00:39:55.689 Uttam Kumaran: It’s just gonna look at it, just make up whatever it sees, yeah, it’s not gonna… which is… it’s just gonna… it’s sort of gonna death spiral into whatever.
310 00:39:55.690 ⇒ 00:40:20.340 Katherine Bayless: Right, but I think we could, we could start with, like, all of the things that, like, we already have the information on, and then, yeah, totally, like, do a working session, brain dump it all in one place, and then I think we can still, lean on the people in the organization to start getting into these QA loops, right? So, like, you know, if we’re working on a table that we know is important, but we don’t know for sure what the definitions are, like, we can loop them in, and maybe we can do it in a clever way, where we’re
311 00:40:20.340 ⇒ 00:40:30.949 Katherine Bayless: collecting the data via, like, an interview, you know, or something like that, and it’s just piping into somewhere we’re grabbing it. But, the other thought I had was…
312 00:40:32.770 ⇒ 00:40:34.360 Katherine Bayless: What was it? Dang it.
313 00:40:34.830 ⇒ 00:40:41.209 Katherine Bayless: Phew. I don’t know, it’ll come back to me. There’s something else you had said, I was like, oh, we could totally lean on that.
314 00:40:42.680 ⇒ 00:40:52.049 Uttam Kumaran: So writing the semantic views, yeah, and then I think the second thing, Amber, is probably, like, I think after… let’s nail the table metadata, and then I think you should just
315 00:40:52.280 ⇒ 00:40:57.920 Uttam Kumaran: Let’s just… I… Try to combine as many semantic views into one as we can.
316 00:40:58.100 ⇒ 00:41:02.889 Uttam Kumaran: like, just overload it and see, like, what you see.
317 00:41:03.510 ⇒ 00:41:08.550 Uttam Kumaran: I think, like, just aim for one, for now, that has more tables.
318 00:41:08.550 ⇒ 00:41:09.220 Katherine Bayless: I don’t.
319 00:41:09.220 ⇒ 00:41:29.020 Uttam Kumaran: mind going over the limit, and then we can ask Awayish… once… I think once you… like, I want to just see… I want you to just see what happens, and then the second piece of that is, like, we can re-architect the tables to have less of them. So… but that’s kind of the order of operations, is I’d rather you stretch the semantic view to go further.
320 00:41:29.170 ⇒ 00:41:30.980 Uttam Kumaran: Just, like, see what happens.
321 00:41:31.320 ⇒ 00:41:34.569 Uttam Kumaran: And, like, I would love you to come back with, like.
322 00:41:34.690 ⇒ 00:41:46.540 Uttam Kumaran: okay, I’m confident we should move to, like, wider tables, and use semantic views, because ultimately, I think your challenge is, like, prove to Catherine that Soaklake Intelligence is, like, actually
323 00:41:46.930 ⇒ 00:41:50.340 Uttam Kumaran: like, It’s faster, smarter, better, you know?
324 00:41:50.520 ⇒ 00:41:51.860 Uttam Kumaran: So, yeah.
325 00:41:51.860 ⇒ 00:41:56.549 Amber Lin: Cool, so I have two work streams I’m comparing. I think the metadata is mostly for the.
326 00:41:56.550 ⇒ 00:41:57.070 Uttam Kumaran: Yes.
327 00:41:57.490 ⇒ 00:41:58.490 Amber Lin: Yeah.
328 00:41:58.760 ⇒ 00:42:01.920 Uttam Kumaran: I think nailed… I think, like, nail the sidebar.
329 00:42:02.270 ⇒ 00:42:07.130 Uttam Kumaran: Yeah, nail the sidebar, make sure the sidebar’s in, like, a good place, with…
330 00:42:07.320 ⇒ 00:42:12.199 Uttam Kumaran: Known limitations, but that it’s actually answering stuff, like, accurately.
331 00:42:12.410 ⇒ 00:42:30.229 Uttam Kumaran: And then the second piece is, like, let’s try to really prove that Snowflake Intelligence is better. If it’s not, then I’m gonna tell them, like, yo, this product, like, what’s going on? So I will also double down on your message, now that I have a really good context. And I’ll do some more digging and ask some people about how they’re tackling this.
332 00:42:31.640 ⇒ 00:42:41.699 Uttam Kumaran: But, like, let’s assume the sidebar is, like, our number one thing to just, like, make sure that’s, like, working really well. And then, in parallel, I haven’t shut down the sandbox, so…
333 00:42:41.700 ⇒ 00:42:42.370 Katherine Bayless: Hmm.
334 00:42:42.370 ⇒ 00:42:47.060 Uttam Kumaran: we could use it, but I guess, Catherine, I could also, maybe at some point this week, like.
335 00:42:47.250 ⇒ 00:42:50.629 Uttam Kumaran: I’ll actually, let me, let me hold back, let me ask Snowflake, is like.
336 00:42:50.830 ⇒ 00:42:58.089 Uttam Kumaran: you really… should we really actually, like, generate another whole instance for this, or, like, what do you suggest? It’s pretty common, I’ve seen this across other clients, they’ve…
337 00:42:58.660 ⇒ 00:43:01.420 Uttam Kumaran: Like, they’ve kept separate instances entirely.
338 00:43:01.730 ⇒ 00:43:05.219 Uttam Kumaran: I… it’s annoying, but we can do that.
339 00:43:05.940 ⇒ 00:43:25.290 Katherine Bayless: Yeah, I mean, yeah, it’s like, I… to me, it might make the most sense, but also if there are other approaches or paradigms to… to handle it. I mean, maybe we could, like, create a data share to ourselves inside the same account? I don’t know, whatever. I remembered what I was gonna say, when you’re talking about the wider tables, I think, interestingly, too.
340 00:43:25.730 ⇒ 00:43:36.169 Katherine Bayless: part of this exercise to figure out, like, you know, how to define the metadata, how to structure the tables, how much a semantic view can really hold, you know, whatever the docs say, setting aside,
341 00:43:36.170 ⇒ 00:43:59.630 Katherine Bayless: I think it will also surface some of the places where, like, we tried to anticipate the few we knew of, but I think we’re finding more of them where it’s, like, the categorizational-type values have changed over the years. Like, we knew those product codes were messy, because, you know, one year it was ampersand, and the next year it was A&D, and it was vehicle tech, and now it’s automotive, and, like, we knew there was a lot of translation happening across years for those codes.
342 00:43:59.630 ⇒ 00:44:03.169 Katherine Bayless: But I think there’s a lot more of those situations out there, like.
343 00:44:03.170 ⇒ 00:44:15.050 Katherine Bayless: part of what I ended up in a fight about the ExpoCAD data with and Claude yesterday was because I guess, at some point, we changed the way we record certain types of meeting room numbers in ExpoCAD, and personally, I’m like.
344 00:44:15.050 ⇒ 00:44:23.810 Katherine Bayless: I don’t think we are reporting on that anywhere. I don’t really care, but… but it’s yet another example of, like, our data’s weird because I was telling it to use…
345 00:44:23.810 ⇒ 00:44:24.500 Katherine Bayless: the, you know.
346 00:44:24.500 ⇒ 00:44:24.880 Uttam Kumaran: Yeah.
347 00:44:24.880 ⇒ 00:44:44.480 Katherine Bayless: Excel sheets to QA the data, and then it was, like, down this rabbit hole of, like, well, but these room codes make no sense, and I’m like, okay, so we probably are going to need a bridge table for room codes at some point. So it’s, like, figuring out what are the things that can go into wide tables, and then how do we support those, like, categorical values and their very squirrely shifts.
348 00:44:46.070 ⇒ 00:44:46.700 Katherine Bayless: Yeah.
349 00:44:46.700 ⇒ 00:44:47.490 Uttam Kumaran: Yeah.
350 00:44:47.600 ⇒ 00:44:48.420 Uttam Kumaran: Yeah.
351 00:44:48.420 ⇒ 00:44:49.550 Katherine Bayless: How.
352 00:44:49.680 ⇒ 00:44:55.029 Uttam Kumaran: Also, like, long… maybe, like, more existential, how married are you to Snowflake, like.
353 00:44:55.370 ⇒ 00:45:07.699 Uttam Kumaran: overall, because some of this, I’m, like, also worried if other applications or services, like, need this context, like, whether we should have… I mean, whether we should start to, like.
354 00:45:07.850 ⇒ 00:45:12.919 Uttam Kumaran: We should centralize this in dbt first, and then it, like, permeates from dbt into Snowflake.
355 00:45:13.200 ⇒ 00:45:17.390 Uttam Kumaran: Similarly, like, the data lake concept is a similar concept, right, where
356 00:45:17.590 ⇒ 00:45:21.599 Uttam Kumaran: In case we need to go back or use that, we have it all there.
357 00:45:21.740 ⇒ 00:45:22.869 Uttam Kumaran: I guess, like.
358 00:45:23.410 ⇒ 00:45:30.810 Uttam Kumaran: my think… and I don’t know, we don’t… maybe we don’t have to make this decision now, but my thinking is, like, that context is use… is gonna be useful
359 00:45:31.610 ⇒ 00:45:37.110 Uttam Kumaran: outside of Snow… potentially outside of Snowflake, maybe the repo is the best place to keep it, yeah.
360 00:45:37.280 ⇒ 00:45:37.800 Katherine Bayless: Yeah.
361 00:45:37.800 ⇒ 00:45:46.419 Amber Lin: If we’re adding metadata, I would really prefer to do it in dbt, because I don’t want me to do it in Snowflake, and it just gets lost, or…
362 00:45:47.190 ⇒ 00:46:00.170 Amber Lin: And I might need… if we’re focusing on sidebar and not semantic views, I need somewhere to do metrics, and that would need a model table, so we would need dbt eventually.
363 00:46:00.170 ⇒ 00:46:08.499 Uttam Kumaran: Yeah, I think you should do it in dbt. I think more of what I… Amber, I need to know is, like, once you put it in dbt, can you… can it… does it just auto-apply to the table?
364 00:46:08.790 ⇒ 00:46:12.229 Uttam Kumaran: Like, I think that’s what you’ll need to explore with the dbt Snowflake integration.
365 00:46:12.230 ⇒ 00:46:12.950 Amber Lin: Hmm, okay.
366 00:46:12.950 ⇒ 00:46:17.499 Uttam Kumaran: Because it needs to live as a snowflake description object.
367 00:46:18.320 ⇒ 00:46:18.820 Katherine Bayless: Yeah.
368 00:46:18.820 ⇒ 00:46:30.710 Uttam Kumaran: So, I don’t know, it’s like, it’s… I would say it’s not an assumption yet that, like, yeah, you could assume that Snowflake has no understanding of, like, the code in the repo, basically, is what I’m saying.
369 00:46:32.340 ⇒ 00:46:39.329 Uttam Kumaran: It would be great, because typically our development process now is, like, we have Snowflake CLI, and we have dbt open, and that, like, loop
370 00:46:39.770 ⇒ 00:46:51.419 Uttam Kumaran: works, but yeah, I think another open item is, like, it doesn’t have much understanding, like, now, of the dbt code itself, outside of, like, that’s just the table DDL, finally.
371 00:46:52.290 ⇒ 00:47:07.390 Katherine Bayless: Which is maybe a good, sort of, additional data point to your question, because I was going to say, I definitely have an answer, like, I… you know, the… one of my, I don’t know, core principles is, like, all of the things I am designing, I want us to be able to cut and run at any.
372 00:47:07.390 ⇒ 00:47:08.820 Uttam Kumaran: Yeah. Okay, okay.
373 00:47:08.820 ⇒ 00:47:15.110 Katherine Bayless: everything’s just moving so fast, right? Like, I don’t think Snowflake’s gonna go anywhere, but if a different product came along, we were like, that.
374 00:47:15.110 ⇒ 00:47:33.659 Katherine Bayless: Right? We’d want to go that direction right away, or if it turns out Mythos comes out and Snowflake’s not secure, we want to run away from it. So yeah, like, I think the data lake and DBT are the core components. It’s like, GitHub, Data Lake, DBT, those are the things I want to invest in, and then, you know, ideally, yeah, they plug into the front-end product, but…
375 00:47:33.660 ⇒ 00:47:36.339 Katherine Bayless: I’m not… I’m not married to Snowflake.
376 00:47:36.650 ⇒ 00:47:41.920 Uttam Kumaran: Okay, great, so that’s a good line in the sand. So, Amber, I think another task we use, like, as part of the
377 00:47:42.160 ⇒ 00:47:44.849 Uttam Kumaran: Table descriptions and column descriptions, like.
378 00:47:45.090 ⇒ 00:47:54.210 Uttam Kumaran: I think there is actually, like, a pretty… there should be some macros or scripts that will help you. As dbt runs, it will actually execute those
379 00:47:54.560 ⇒ 00:47:57.959 Uttam Kumaran: table descriptions, like, I’m sure when you’re writing
380 00:47:58.180 ⇒ 00:48:02.410 Uttam Kumaran: column and table descriptions in YAML and dbt, there’s a way to just have it go apply.
381 00:48:02.580 ⇒ 00:48:08.189 Uttam Kumaran: So if you could take that on, too. That way, that becomes a source of truth for descriptions of stuff.
382 00:48:08.310 ⇒ 00:48:12.640 Uttam Kumaran: And ultimately, we can source that from… we can move that to source from, like.
383 00:48:12.850 ⇒ 00:48:16.630 Uttam Kumaran: Catalog or something where people can edit it over time, you know?
384 00:48:16.830 ⇒ 00:48:18.749 Katherine Bayless: yeah.
385 00:48:19.000 ⇒ 00:48:23.340 Katherine Bayless: Yeah, that will be… I don’t think that’s a V0 problem.
386 00:48:23.340 ⇒ 00:48:23.860 Uttam Kumaran: Yes.
387 00:48:23.860 ⇒ 00:48:30.029 Katherine Bayless: Like, the… what was the rule when this data happened versus what is it, like, now?
388 00:48:30.160 ⇒ 00:48:41.230 Katherine Bayless: that’ll eventually matter, but, I mean, we can start with current definitions, and current metrics, and when we get into, like, weird historical situations, we can tackle that when it happens.
389 00:48:41.560 ⇒ 00:48:42.240 Uttam Kumaran: Okay.
390 00:48:42.300 ⇒ 00:48:45.979 Katherine Bayless: But I’m sure there are definitions that have shifted over the years.
391 00:48:47.140 ⇒ 00:48:52.940 Uttam Kumaran: Yeah, and so part of this is, like, if something shifts and someone notices it, can they just go somewhere and edit it? It’s like a UI.
392 00:48:53.230 ⇒ 00:48:58.930 Katherine Bayless: But it’s also, like, how do we have them edit it, but not lose the, like, okay, but data collected prior to this date.
393 00:48:58.930 ⇒ 00:49:00.520 Uttam Kumaran: Yes, yeah, yeah.
394 00:49:00.520 ⇒ 00:49:01.370 Katherine Bayless: Yeah, right, I know.
395 00:49:01.370 ⇒ 00:49:03.290 Uttam Kumaran: Oh, yeah, nightmare. Nightmare.
396 00:49:03.290 ⇒ 00:49:04.519 Katherine Bayless: We’ll solve that later.
397 00:49:04.520 ⇒ 00:49:07.640 Uttam Kumaran: Nightmare.
398 00:49:07.640 ⇒ 00:49:14.979 Katherine Bayless: It’s funny, at my old place, so we had, like, the two, startups that we were running, and the…
399 00:49:15.390 ⇒ 00:49:31.099 Katherine Bayless: all the original CRM and then the two startups, inevitably, we would come back to these, like, conversations about business rules and, like, how do you really get your arms around them? The only project we had the entire time I was there that we never managed to get launched was a rules engine. Like, we really wanted to
400 00:49:31.100 ⇒ 00:49:37.709 Katherine Bayless: Yeah. …act thing, like a state machine for business logic over time, and yeah, we never…
401 00:49:37.710 ⇒ 00:49:50.329 Uttam Kumaran: That was the company I led… I was… I was doing product at. Oh, yeah? Before Brainforge. Yeah, I led… we basically built… I was like, that… that’s… yeah, I, like, led the product development for basically, like, a no-code semantic layer builder.
402 00:49:50.630 ⇒ 00:49:56.360 Uttam Kumaran: tool that you could build these rules, and, like, we… I had to… you sort of learned everything about, like.
403 00:49:56.560 ⇒ 00:50:05.159 Uttam Kumaran: every version of dimensions and metrics, we have these, like, deep conversations about, like, what’s a column, what’s a metric, and, like, how do we… yeah, but…
404 00:50:05.830 ⇒ 00:50:10.020 Uttam Kumaran: No longer. Yeah, we.
405 00:50:10.020 ⇒ 00:50:29.089 Katherine Bayless: We managed to get our arms around it for accreditation. That was, like, one of the big use cases we had initially, was, like, is somebody’s degree accredited? It depends. When did they go? What were the rules for accreditation at that time? Because there’s grace periods. If you get accredited halfway through your degree, we give you the whole thing. If it happens 3 months later, it’s too bad, you should go back to school. And so yeah, this, like.
406 00:50:29.090 ⇒ 00:50:34.950 Katherine Bayless: business rules over time state machine thing. Great idea! Difficult to execute.
407 00:50:34.950 ⇒ 00:50:36.260 Uttam Kumaran: Yes, yes.
408 00:50:36.260 ⇒ 00:50:36.800 Katherine Bayless: Yeah.
409 00:50:37.720 ⇒ 00:50:40.210 Katherine Bayless: But yeah, I like… I like this…
410 00:50:40.300 ⇒ 00:50:59.629 Katherine Bayless: path of table metadata, I definitely think we can lean on the users to help define the stuff that we don’t know much about, and then kind of figuring out how we want to handle the, like, bridge tables for lookups for things that have changed over the years, code-wise. And then, yeah, like, setting up
411 00:50:59.630 ⇒ 00:51:04.109 Katherine Bayless: Either a, you know, different instance or some other method where we can
412 00:51:04.110 ⇒ 00:51:15.699 Katherine Bayless: because I really… I don’t necessarily want to abandon the semantic views. I think the idea of putting context around your data is something that is getting out there for a very, very long time. I just… I don’t know…
413 00:51:15.870 ⇒ 00:51:17.759 Katherine Bayless: If we wanna, like…
414 00:51:18.150 ⇒ 00:51:26.110 Katherine Bayless: wait and see what Snowflake changes, wait and see how we can, you know, what we can learn on our side, but, like, yeah, like…
415 00:51:26.110 ⇒ 00:51:26.460 Uttam Kumaran: Yeah.
416 00:51:26.820 ⇒ 00:51:30.369 Amber Lin: Let’s… let’s try to do the big one. It’s essentially the same as metadata.
417 00:51:30.370 ⇒ 00:51:32.299 Katherine Bayless: Yeah, that’s rude, yeah, yeah, yeah, that too, yeah.
418 00:51:32.300 ⇒ 00:51:50.099 Amber Lin: derived metrics, we can just add those in the table, in a new table, and then as long as we have a file to define the join relationships, between table, ideally between everything that has relationships, it should… it should just perform better.
419 00:51:50.190 ⇒ 00:51:53.180 Amber Lin: Actually,
420 00:51:53.460 ⇒ 00:51:57.940 Katherine Bayless: I realize that’s a good question. When you say join, I…
421 00:51:58.380 ⇒ 00:52:09.919 Katherine Bayless: So, like, none of the systems share keys for the most part, so we would basically be saying, like, maybe join on email, maybe join on this, maybe join on that, or is it, like, you get one?
422 00:52:10.340 ⇒ 00:52:25.409 Amber Lin: Some of the joints, I think, ways where the team has been looking for, I think, especially between membership and between attendance, or between exhibitor and attendance, we’ve had the identity stitching tables that were used.
423 00:52:26.540 ⇒ 00:52:28.889 Amber Lin: Within schemas, of course, they join.
424 00:52:29.170 ⇒ 00:52:29.650 Katherine Bayless: Yeah.
425 00:52:29.650 ⇒ 00:52:39.950 Amber Lin: And that’s mostly what I’ve been dealing with, with dismantic views. So when we go cross tables, even if we use a sidebar, that’s gonna be an issue we’ll need to tackle.
426 00:52:40.340 ⇒ 00:52:59.719 Katherine Bayless: Yeah, yeah, it’s like the Cortex Code agent manages to get by, and I think it typically leans on email, which is basically what we use, but yeah, like, it’s gonna get, it’ll get interesting. So I like, sort of, the idea of testing out and seeing if we do try the giant semantic view, and the
427 00:52:59.790 ⇒ 00:53:03.900 Katherine Bayless: you know, scrappy join logic that we have. How good does it do?
428 00:53:04.220 ⇒ 00:53:04.900 Amber Lin: Yeah. Central.
429 00:53:05.020 ⇒ 00:53:24.129 Amber Lin: I’m a little bit concerned when it comes to things that we have stitched already, like for membership. The answer was much… a lot better when we tried to use the identity stitching tables, but if it’s a massive semantic view and it roams free, I don’t know which table it’ll pick.
430 00:53:25.010 ⇒ 00:53:28.840 Amber Lin: Most cases, a smaller semantic view serves this purpose.
431 00:53:29.260 ⇒ 00:53:30.740 Katherine Bayless: Yeah, yeah.
432 00:53:30.920 ⇒ 00:53:55.670 Katherine Bayless: But I think, too, one of the pieces I want to get off my plate, today, tomorrow, so I’m almost done with the ExpoCAD stuff, and then the next thing is that entity identification service that’s actually, like, the API. And then that way, instead of us… I mean, we can keep the tables, or we can maybe replace them with the DynamoDB tables that are the backend for the API, but, like, we would be able to also use that as, like, a lookup just in time.
433 00:53:55.760 ⇒ 00:53:56.720 Katherine Bayless: Hmm.
434 00:53:56.920 ⇒ 00:53:57.510 Katherine Bayless: Yeah.
435 00:53:57.800 ⇒ 00:54:05.020 Amber Lin: Yeah, and I just thought of this. The big semantic view can just be our generic agent, like.
436 00:54:05.120 ⇒ 00:54:23.869 Amber Lin: And then we can keep the smaller views for more accurate things that we’ve verified, but if we wanted to… want the agent to talk to everything, then a semantic view that covers everything but may draw in the wrong thing will just be the… will just be problem arts that it has access to.
437 00:54:24.280 ⇒ 00:54:26.080 Katherine Bayless: Yeah, yeah.
438 00:54:26.250 ⇒ 00:54:35.100 Katherine Bayless: I know it’s… I really… it’s like, it’s definitely our data that is the challenge, but… Yeah.
439 00:54:36.380 ⇒ 00:55:01.220 Katherine Bayless: Because, like, even bringing in the Zoom data, right? Like, we’re going to want to join that on, yes, the individual people who attended the Zooms, are they in remembers, but then we’re also going to want to join it on the companies they work for, is that company a member? And then we’re going to join it on the topic of the webinar, right? Like, a digital health webinar and digital health session at CES, people are going to want to see, right? Like, overlaps and convergence there. Like, we’re going to want to join the webinars based on
440 00:55:01.220 ⇒ 00:55:03.660 Katherine Bayless: Committee names, if it was a committee call, like.
441 00:55:03.670 ⇒ 00:55:08.519 Katherine Bayless: Just like, yeah, like, there’s so many different ways this comes together. I don’t know.
442 00:55:11.430 ⇒ 00:55:12.310 Amber Lin: I see.
443 00:55:12.500 ⇒ 00:55:26.179 Katherine Bayless: And it’s almost like you kind of wonder, like, maybe semantic views, and I don’t mean, like, how we interact with them now, but, like, what they might evolve into is, like, it’s almost like the joins are the part that needs to go away, right? Because, like, it’s not…
444 00:55:26.560 ⇒ 00:55:28.360 Katherine Bayless: It’s not as challenging to…
445 00:55:29.320 ⇒ 00:55:36.340 Katherine Bayless: Like, all of the, like, metadata and explanations on this stuff, but, like, the joins are…
446 00:55:36.460 ⇒ 00:55:40.209 Katherine Bayless: so much more nebulous. It really depends on the business case, you know?
447 00:55:40.510 ⇒ 00:55:41.309 Amber Lin: Huh.
448 00:55:42.020 ⇒ 00:55:48.010 Amber Lin: Wait, do you mean that the joins in the semantic views right now, or…
449 00:55:48.230 ⇒ 00:56:00.810 Katherine Bayless: We’re not necessarily, like, in the semantic views as they exist today, but, like, the way we use our data, right? Like, if somebody asks, like, you know, what, what’s the attendance rate on the AV board?
450 00:56:00.880 ⇒ 00:56:14.309 Katherine Bayless: Right? Like, they’re gonna need to know on Remembers, I have to go find the AV board, find their people, then I need to go over to Zoom, but I’m not looking for those people, I’m looking for Zoom calls that were for the AV board.
451 00:56:14.310 ⇒ 00:56:22.649 Katherine Bayless: You know, right? But then, most of the time, we wouldn’t necessarily need to go at a committee-level join, but, like, I don’t know, it’s like, it’s just, it’s almost…
452 00:56:22.900 ⇒ 00:56:32.810 Katherine Bayless: It feels like there’s so many different ways that people might need to connect the data that aren’t always explicit join keys. That’s where I’m like, I wonder if future versions of semantic keys will be more flexible.
453 00:56:32.810 ⇒ 00:56:35.669 Amber Lin: It’s… I think it will be…
454 00:56:35.910 ⇒ 00:56:48.590 Amber Lin: I like defining the dimensions, defining things as, hey, this is the main key you should join on, this is the key you would use in this case, and just adding context in terms of
455 00:56:48.750 ⇒ 00:56:51.900 Amber Lin: The join key, maybe?
456 00:56:51.900 ⇒ 00:56:52.620 Katherine Bayless: Everything?
457 00:56:52.620 ⇒ 00:57:10.209 Amber Lin: add a file to say, hey, maybe in Cortex search or something, like, hey, this is the relationships file of… here are potential joins that work, here are… if you join on this, you will miss this percentage of unverified,
458 00:57:11.740 ⇒ 00:57:17.979 Amber Lin: but you can do it if it’s last resort, like, I think we can still specify those things right now.
459 00:57:18.510 ⇒ 00:57:27.779 Katherine Bayless: Yeah, well, I mean, it’s like… and or, like, I could see where it’d be like, in this semantic view, you’re like, do all of that, but then also kind of leaving it open-ended to say, like.
460 00:57:27.780 ⇒ 00:57:46.049 Katherine Bayless: But also, like, use your own reasoning based on what the user asked, right? Like, that’s the kind of future state that I think might wind up existing. Because, like, we wouldn’t have a key, necessarily, we’re just doing a, you know, a semantic match on… or a fuzzy match, rather, on names that may or may not be the same in both places, like…
461 00:57:46.050 ⇒ 00:57:52.560 Katherine Bayless: Reasoning is really how we’re joining these things, I guess. Which is, slippery, yeah.
462 00:57:52.900 ⇒ 00:57:58.410 Amber Lin: Yeah, I mean, when it’s not using semantic use, that’s… that’s how it figures things out.
463 00:57:58.410 ⇒ 00:57:59.229 Katherine Bayless: Yeah, exactly.
464 00:57:59.230 ⇒ 00:58:01.220 Amber Lin: just figured this out.
465 00:58:01.220 ⇒ 00:58:01.980 Katherine Bayless: Yeah.
466 00:58:02.140 ⇒ 00:58:03.610 Amber Lin: That’s what it does.
467 00:58:03.950 ⇒ 00:58:16.349 Katherine Bayless: Yeah. So it’s like, that’s what I wish we could say, is like, here’s all the stuff that’s in these tables, and here’s what it means, and here’s the business rules, but, like, join based on reasoning. Don’t worry about what exactly, like, may or may not key between the tables.
468 00:58:16.350 ⇒ 00:58:16.920 Amber Lin: Hmm.
469 00:58:17.150 ⇒ 00:58:17.810 Katherine Bayless: Right?
470 00:58:17.960 ⇒ 00:58:21.160 Katherine Bayless: I mean, I know that’s not possible at the moment, I just, like…
471 00:58:21.780 ⇒ 00:58:26.180 Katherine Bayless: Feels like it would be the thing that would unblock some of the friction.
472 00:58:28.510 ⇒ 00:58:29.899 Amber Lin: Yeah, I can… I can…
473 00:58:30.010 ⇒ 00:58:35.300 Amber Lin: I can see if… if I can do that in the bigger semantic views.
474 00:58:35.460 ⇒ 00:58:51.180 Katherine Bayless: I mean, I’m, like, blue-sky-ing. I’m like, just… because I really, I can’t imagine we’re the only people losing Snowflake this way and running into, like, the challenges on this side, and so I’m like, you know, I just… I start to wonder, like, how will this problem get solved at scale?
475 00:58:51.840 ⇒ 00:58:52.750 Amber Lin: Yeah.
476 00:58:52.910 ⇒ 00:59:10.129 Amber Lin: I was wondering that as well as why can’t we query propmars directly in the agent? Like, why limit that ability to just the sidebar when you apparently have a new interface that you’re trying to do?
477 00:59:10.190 ⇒ 00:59:16.140 Amber Lin: Maybe they’re trying to separate the two products of sidebar versus… Intelligence.
478 00:59:16.460 ⇒ 00:59:17.470 Amber Lin: I don’t know.
479 00:59:17.960 ⇒ 00:59:34.410 Katherine Bayless: I… my suspicion is that, like, because they had announced that whole snowwork thing back in, like, February or March, and, like, it hasn’t rolled out to us yet, but, like, I think Snowflake Intelligence will become snowwork, and I think they will intend for that to be the, like, end user.
480 00:59:34.410 ⇒ 00:59:35.110 Amber Lin: they’re facing.
481 00:59:35.110 ⇒ 00:59:54.660 Katherine Bayless: side, yeah, exactly. And then, like, the Cortex code sidebar chat is, like, for your developer who randomly happens to be in there at that moment. Yeah, because, like, honestly, I use, like, most of my Snowflake work I do with Cloud Code and the CLI, and so it’s not even using, like, the analyst or the agent at all, it’s just…
482 00:59:54.730 ⇒ 01:00:03.200 Katherine Bayless: using its own brain to write the sequel. But but yeah, I agree. Like, I think Snowflake is gonna start to, like, diverge.
483 01:00:04.850 ⇒ 01:00:12.699 Amber Lin: Yeah, it’s… in that case, we can try the custom tool. It will be, essentially.
484 01:00:12.750 ⇒ 01:00:27.519 Amber Lin: calling an LLM and say, hey, this is a text-to-SQL tool, you have these certain boundaries, and this is what you’ll do. What do you think about that? There are some downsides in terms of, like, grant permissions and roles and all that.
485 01:00:28.430 ⇒ 01:00:33.440 Katherine Bayless: I mean, I… yeah, I’m inclined to agree with you, I feel like…
486 01:00:33.680 ⇒ 01:00:42.700 Katherine Bayless: a good thing to test if none of the other ideas we’ve had wind up working out, but hopefully, like… because, yeah, that does feel like a…
487 01:00:44.040 ⇒ 01:00:48.620 Katherine Bayless: it feels more band-aid-y, and more, like, you know…
488 01:00:48.620 ⇒ 01:00:50.169 Amber Lin: Duplicative, because then we’ll have to…
489 01:00:50.170 ⇒ 01:00:50.910 Katherine Bayless: Okay.
490 01:00:50.910 ⇒ 01:00:59.929 Amber Lin: the text to SQL as things improve, and it’ll rely on the metadata anyway, so maybe we’ll see how it goes.
491 01:01:00.450 ⇒ 01:01:10.739 Katherine Bayless: Yeah, like, I really… I think the metadata is the way to start, and I think, like, from there, we’ll start to see the agent do a little bit better, generally, and…
492 01:01:10.980 ⇒ 01:01:11.840 Katherine Bayless: Yeah.
493 01:01:12.550 ⇒ 01:01:13.340 Katherine Bayless: Yeah.
494 01:01:14.300 ⇒ 01:01:22.860 Amber Lin: Alright. I did add the remaining semantic views from QuadMart, so right now it should cover…
495 01:01:23.120 ⇒ 01:01:32.810 Amber Lin: all of the tables in PropMarch, but they’re not, like, refined. I don’t know what the numbers should be.
496 01:01:33.840 ⇒ 01:01:55.170 Katherine Bayless: I mean, that’s true, I mean, some of these we don’t necessarily… I mean, I… like, for Shopify, for example, like, people have never been able to really report on this data before, so it’s not like there is, like, a, you know, official number. Like, I could probably talk to finance and be like, what did we file in our taxes? Right? But yeah, like, there’s gonna be some of this stuff that we’re like, I don’t know, hope it’s right.
497 01:01:55.730 ⇒ 01:02:12.099 Amber Lin: Hmm, I see. I think the one-on-ones will be helpful then, because I do want to talk to someone about, hey, is this correct? Because I don’t even know what all of the… necessarily what all of the fields are. Right. Right now…
498 01:02:12.370 ⇒ 01:02:20.880 Amber Lin: Everything is in… they’re in… They’re frond, so we have…
499 01:02:23.980 ⇒ 01:02:40.249 Amber Lin: this is an older one, but we have attendance, sessions, event, exhibitor, and then we have the Innovation Awards, SFMC, and Shopify, which are the new ones I just added. How would you like to do…
500 01:02:40.360 ⇒ 01:02:50.129 Amber Lin: the rest of the working session. Would you like to work on anything particular? I can also go ahead and try what we proposed today. I can make, like, a
501 01:02:50.520 ⇒ 01:02:54.440 Amber Lin: overall semantic view that covers everything in prop marts.
502 01:02:54.700 ⇒ 01:03:06.120 Amber Lin: And then I can try to work on one of the table’s metadata. What would you… what’s a better use of your time? Because we can always meet later today. My calendar’s fully open.
503 01:03:06.510 ⇒ 01:03:16.230 Katherine Bayless: That’s true, actually, yeah, the rest of the day is open for me, too, and… and I do, I want to, like, whatever is the most useful,
504 01:03:18.110 ⇒ 01:03:25.269 Katherine Bayless: my… my brain gravitates toward, like, I can probably, like, brain dump metadata thoughts?
505 01:03:25.540 ⇒ 01:03:28.859 Katherine Bayless: But if you wanted to, like, I guess…
506 01:03:29.340 ⇒ 01:03:41.910 Katherine Bayless: Depending on how long it would take for you to try the, like, one big view thing, like, if you want to, like, kick the tires on that, and then ping me in an hour or two, and then we can hop back on and look at it, or if it’s like, that’s gonna take two days, ma’am. That’s fine too.
507 01:03:41.910 ⇒ 01:03:53.939 Amber Lin: I don’t think so. I have… I have all the files, so I’ll probably just generate one. The cross schema joins would be wonky, and probably I’ll have to talk to you about that anyways.
508 01:03:54.050 ⇒ 01:03:58.630 Amber Lin: But I’ll have a semantic view of
509 01:03:59.220 ⇒ 01:04:03.220 Amber Lin: different schemas with some unrelated tables, but then I can just…
510 01:04:03.360 ⇒ 01:04:06.180 Amber Lin: Like, bare minimum, just dump all the tables in one.
511 01:04:07.010 ⇒ 01:04:14.730 Katherine Bayless: Yeah, I mean, it’s up to you, right? Like, I think, I don’t mind whichever way you want to go, just, like, whatever is the most,
512 01:04:15.020 ⇒ 01:04:25.380 Katherine Bayless: whatever’s actually the most useful, because I do sometimes feel like I’m, like, the chaos agent that’s just like, hi, here’s a whole bunch of trouble. And so, you know, put me in the box that makes sense, but…
513 01:04:25.380 ⇒ 01:04:43.759 Amber Lin: Let me… let me do the big semantic view real quick, probably an hour or two, and test it a little bit. I’ll send you stuff async in the channel, and then I think the metadata will probably take a lot more, like, back and forth and working on them.
514 01:04:43.820 ⇒ 01:04:54.599 Amber Lin: Is there a table you would recommend to start with the metadata? I’ll probably do, like, registration and attendance, because that’s the one thing I actually know a little bit more about.
515 01:04:54.850 ⇒ 01:05:01.660 Katherine Bayless: Right, yeah, no, that’s probably a good one to start with, honestly, because same, right? I mean, yeah.
516 01:05:01.660 ⇒ 01:05:15.239 Katherine Bayless: Yeah, I would say either the registration attendance stuff, or we could go the exhibitor direction, since my head’s been in that space for a little bit now, but, yeah, I think both of those would be a good place to start, where I would feel fairly confident explaining the fields.
517 01:05:15.240 ⇒ 01:05:17.479 Katherine Bayless: To your point, some of the other stuff, like…
518 01:05:17.510 ⇒ 01:05:19.739 Katherine Bayless: I don’t know, and I’m not sure who does.
519 01:05:20.400 ⇒ 01:05:30.750 Amber Lin: Cool, okay. For the big overall semantic views, maybe I should also try, if we have the fatter tables, to.
520 01:05:31.410 ⇒ 01:05:34.710 Amber Lin: Do one from that and compare.
521 01:05:34.970 ⇒ 01:05:35.540 Katherine Bayless: Yeah…
522 01:05:35.540 ⇒ 01:05:41.449 Amber Lin: I don’t know how many we have. I know we have it for CES.
523 01:05:41.780 ⇒ 01:05:53.209 Katherine Bayless: Right. I was gonna say, that’s the main one that we would have the, like, fatter table for, would be the registration, export from CES, the old exhibitor report, same thing.
524 01:05:53.430 ⇒ 01:06:01.749 Katherine Bayless: I… I think the Innovation Awards… I’m not… to be totally honest, I’m not sure what
525 01:06:01.750 ⇒ 01:06:23.439 Katherine Bayless: shape they ended up in going through dbt, because Kyle’s kind of taken the lead on that, but at the very least, when you export them from the platform, they are massively wide table. I mean, like, I think they’re 100 columns or more kind of thing, right? Because it’s every application question, and so the Innovation Awards would be another, like, big, fat, wide table, I think.
526 01:06:23.700 ⇒ 01:06:26.860 Katherine Bayless: Okay. Maybe pre-DBT, potentially post.
527 01:06:27.830 ⇒ 01:06:32.900 Amber Lin: Mmm, okay, I’ll try to find those tables,
528 01:06:33.520 ⇒ 01:06:43.139 Amber Lin: And then I’ll also make one from those, just a rough one. But I know, like, maybe definitions or things have changed when we transformed them, so we’ll see.
529 01:06:43.140 ⇒ 01:06:50.379 Katherine Bayless: True, yeah, yeah. And honestly, the Innovation Awards is probably some of the messiest data, just because it’s…
530 01:06:51.090 ⇒ 01:07:01.960 Katherine Bayless: Yeah, it’s one of those programs that has not had a system, necessarily, right? I mean, it’s always had something, but we tend to, like, hate it and try something new every year, and so it’s just… yeah.
531 01:07:01.960 ⇒ 01:07:02.830 Amber Lin: A little…
532 01:07:02.830 ⇒ 01:07:08.740 Katherine Bayless: A little chaotic. And that’s what I told everybody, I was like, we’ve been doing disposable software this whole time, we’ve just been buying it instead of building it.
533 01:07:09.000 ⇒ 01:07:11.719 Amber Lin: I see.
534 01:07:12.020 ⇒ 01:07:23.700 Katherine Bayless: Yeah, like, it’s, it’s an interesting, interesting environment here. Like, people are lovely. Tech and data have been chaos, but, slowly raining it in.
535 01:07:24.130 ⇒ 01:07:39.239 Amber Lin: Okay, I think we can regroup in an hour or hour and a half. I’ll put a… I’ll put a placeholder then, and then, like, if we need to push it back, or if we can meet early, I’ll message you on Slack.
536 01:07:39.610 ⇒ 01:07:41.929 Katherine Bayless: Okay, cool, cool. I’ll just be here, so let me know.
537 01:07:42.260 ⇒ 01:07:43.960 Amber Lin: Thanks, talk to you later.
538 01:07:43.960 ⇒ 01:07:44.910 Katherine Bayless: Thanks, Amber. Bye.
539 01:07:44.910 ⇒ 01:07:45.450 Amber Lin: Bye.