WEBVTT 1 00:00:01.560 --> 00:00:05.799 Are you looking for a good union job? The Inland Empires, 2 00:00:05.879 --> 00:00:10.000 fourteen thousand members strong Teamsters Local nineteen thirty two has 3 00:00:10.160 --> 00:00:13.919 opened a training center to get working people trained and 4 00:00:14.039 --> 00:00:17.640 placed in open positions in public service, clerical work, and 5 00:00:17.800 --> 00:00:21.320 in jobs in the logistics industry. This is a new 6 00:00:21.399 --> 00:00:26.000 opportunity to advance your career and raise standards across the region. 7 00:00:26.559 --> 00:00:31.120 Visit nineteen thirty two Training Center dot org to enroll today. 8 00:00:31.600 --> 00:00:35.280 That's nineteen thirty two Training Center dot org. 9 00:00:37.079 --> 00:00:41.039 Listina KCAA Lewelinda at one O six point five FMK 10 00:00:41.119 --> 00:00:42.880 two ninety three CF Brino Valley. 11 00:00:43.600 --> 00:00:47.119 The information economy as a rod. The world is teeming 12 00:00:47.240 --> 00:00:50.359 with innovation as new business models reinvent. 13 00:00:50.079 --> 00:00:51.439 Every industry industry. 14 00:00:51.520 --> 00:00:55.880 Inside Analysis is your source of information and insights about 15 00:00:55.880 --> 00:00:58.399 how to make the most of this exciting new era. 16 00:00:58.679 --> 00:01:02.920 Learn more at inside analysis dot comsideanalysis dot com. And 17 00:01:03.000 --> 00:01:05.760 now here's your host, Eric Kavanaugh. 18 00:01:10.400 --> 00:01:13.760 Hello, and welcome to Inside Analysis. I'm your host and 19 00:01:13.920 --> 00:01:19.640 Yell LeBlanc. On this episode, Eric Kavanaugh is interviewing Alex Galago, 20 00:01:20.000 --> 00:01:24.120 who's the CEO of Red Panda Data. Together, they will 21 00:01:24.159 --> 00:01:28.319 talk through the new landscape of LMS and how the 22 00:01:28.400 --> 00:01:33.560 innovation of AI is affecting companies all around. Stay tuned, 23 00:01:33.599 --> 00:01:35.560 you're listening to inside Analysis. 24 00:01:35.879 --> 00:01:38.319 I'd like to dig deeper into Red Panda. 25 00:01:38.359 --> 00:01:40.719 I mean, I know you've you've basically rewritten the whole 26 00:01:40.719 --> 00:01:44.680 platform to do streaming, to do it efficiently, because Java 27 00:01:44.760 --> 00:01:47.079 is not very efficient, right, So you said, well, just 28 00:01:47.120 --> 00:01:49.560 go to the drawing board and hack this thing out. 29 00:01:50.159 --> 00:01:51.359 How long did it take you to do that? 30 00:01:52.319 --> 00:01:56.000 Beca In many ways it feels. 31 00:01:57.200 --> 00:02:00.359 It felt to me like it was a natural evolution of, 32 00:02:01.040 --> 00:02:05.280 you know, decades of work. While I was in school, 33 00:02:05.439 --> 00:02:10.360 I took a job at a Forks trading company and 34 00:02:10.479 --> 00:02:12.400 I really got sort of introduced to a bunch of 35 00:02:12.400 --> 00:02:18.080 low latency things and how do markets work? And I 36 00:02:18.159 --> 00:02:21.960 ended up writing some some trading algorithms in a weird 37 00:02:22.039 --> 00:02:26.120 language called MQL four. It's like this weird obscure language, 38 00:02:26.280 --> 00:02:30.080 kind of Jenkie, to be honest, very Python like, but 39 00:02:30.479 --> 00:02:35.680 not at school. And then I wrote it at a 40 00:02:35.719 --> 00:02:39.439 bunch of trading apps on this thing called pips, and 41 00:02:39.439 --> 00:02:41.680 a pip is like whatever, like one hundred dollars or 42 00:02:41.759 --> 00:02:44.039 something like that, just like some really large spread. But 43 00:02:44.080 --> 00:02:46.360 you want to deal in terms of bacurns. This long 44 00:02:46.400 --> 00:02:50.439 story sort, that's where I really started to kind of 45 00:02:50.520 --> 00:02:53.319 lean into things that I found really fun, which were 46 00:02:53.759 --> 00:02:56.960 like low latency, high performance things. And so then I 47 00:02:57.000 --> 00:02:59.639 went to work for an Attech. I guess when I graduated, 48 00:02:59.719 --> 00:03:02.960 I worked in an embedded database and I worked on 49 00:03:04.520 --> 00:03:07.080 the greed this is for the BlackBerry phone. We wrote 50 00:03:07.080 --> 00:03:11.080 this like c parsing code where we were pulling this 51 00:03:11.159 --> 00:03:14.639 like three dimensional database into a series of doubles because 52 00:03:14.639 --> 00:03:18.879 it was really efficient to encode for I think open VMS, 53 00:03:18.919 --> 00:03:23.120 and so we wrote the color coding on the S 54 00:03:23.159 --> 00:03:27.560 and P five hundred table with like dynamic rendering. That 55 00:03:27.639 --> 00:03:30.319 was really cool and so long story short, when I 56 00:03:30.360 --> 00:03:34.199 think about the evolution of red Panda, it felt to 57 00:03:34.240 --> 00:03:36.439 me like the natural evolution of the ideas that I 58 00:03:36.479 --> 00:03:40.159 had worked on, which went to be an Attech in 59 00:03:40.240 --> 00:03:43.319 New York for a company called Yildmo, and then for 60 00:03:43.360 --> 00:03:46.080 a stream processing company that I wrote the tech for 61 00:03:48.039 --> 00:03:52.479 in twenty fourteen through twenty sixteen that I sold to Akamai. 62 00:03:52.960 --> 00:03:56.400 And so anyways, in many ways, it was like, why 63 00:03:56.439 --> 00:03:58.599 is the world so complicated? Why does it have to 64 00:03:58.599 --> 00:04:01.000 be so hard? And everyone that has been on call 65 00:04:01.560 --> 00:04:03.599 and you know, usually your page I goes up at 66 00:04:03.599 --> 00:04:06.199 three am. I don't know why, but it's like somewhere 67 00:04:06.240 --> 00:04:08.319 around that time. It's like though, it's when you're like 68 00:04:08.400 --> 00:04:11.280 writing your deep sleep and you get paid and then 69 00:04:11.319 --> 00:04:13.360 you wake up and you're like, I don't understand why 70 00:04:13.360 --> 00:04:14.479 this is so complicated. 71 00:04:15.560 --> 00:04:15.879 And so. 72 00:04:17.360 --> 00:04:19.639 It was like, well, why couldn't we do better? Like 73 00:04:19.720 --> 00:04:22.120 what's what's the mechanical reason? Like, okay, let's remove all 74 00:04:22.160 --> 00:04:26.319 the hype. I just want to understand fundamentally, what is 75 00:04:26.360 --> 00:04:28.240 the job that this software is doing that is so 76 00:04:28.480 --> 00:04:30.959 hard that requires it to be so complicated. It turned 77 00:04:30.959 --> 00:04:34.279 out not to be that difficult, frankly, and so I 78 00:04:34.319 --> 00:04:40.879 wrote something from a scratch really for me, and I 79 00:04:40.959 --> 00:04:47.680 left Akamai I think December twenty eighteen somewhere around there, 80 00:04:47.879 --> 00:04:51.000 like basically, and I started writing Red Panda in January. 81 00:04:51.560 --> 00:04:54.480 But I would say I had a like a pretty 82 00:04:54.560 --> 00:04:56.399 clear goal and so it was just a matter of 83 00:04:56.480 --> 00:04:59.920 writing the code. And so that part didn't take that long, 84 00:05:00.000 --> 00:05:02.560 maybe like three months to have a prototype, a couple 85 00:05:02.600 --> 00:05:06.480 of months to have something well that that works interesting. 86 00:05:07.160 --> 00:05:09.160 And of course the world has just sped up in 87 00:05:09.199 --> 00:05:12.399 all directions around us. Now, so we're all kind of 88 00:05:12.519 --> 00:05:16.240 riding this new wave. And you know, when I think 89 00:05:16.279 --> 00:05:19.279 about streaming, I've always i mean from the earliest days, 90 00:05:19.319 --> 00:05:21.199 I was like, oh my goodness, if you do this right, 91 00:05:21.959 --> 00:05:26.120 you can just you can supersede ancient architectures or what 92 00:05:26.160 --> 00:05:28.560 should be viewed as ancient architectures. But you have to 93 00:05:28.560 --> 00:05:30.680 do the whole thing. It's like when you try to 94 00:05:30.720 --> 00:05:34.480 fuse these two worlds together, that's where the impedance mismatch 95 00:05:34.920 --> 00:05:37.120 kicks into high gear and you've got to figure out 96 00:05:37.120 --> 00:05:37.399 some way. 97 00:05:37.439 --> 00:05:38.079 I mean, you know, I. 98 00:05:38.079 --> 00:05:41.720 Heard stories of at Vertica where they were taking all 99 00:05:41.800 --> 00:05:45.040 kinds of streaming dages just like filling it out into 100 00:05:45.199 --> 00:05:47.839 you know, a traditional database, and you know it's a 101 00:05:47.879 --> 00:05:50.439 column oriented database of course, but still it's like, Okay, 102 00:05:50.600 --> 00:05:52.279 you know, do you need all that? Like is that 103 00:05:52.519 --> 00:05:55.680 really really necessary? And the short answer, of course is no, 104 00:05:55.759 --> 00:05:58.560 not really. But you know, for for like net new 105 00:05:58.680 --> 00:06:01.600 use cases, I think you guys would be a dream, right, 106 00:06:01.680 --> 00:06:04.680 But in traditional architectures do you find that's kind of 107 00:06:04.879 --> 00:06:08.560 a hindrance is that the behemistic are the big fortune 108 00:06:08.600 --> 00:06:11.519 two thousand companies They just you know, they're like, oh 109 00:06:11.560 --> 00:06:13.519 my god, like can we even handle this is that? 110 00:06:14.240 --> 00:06:17.040 Is that still a hurdle or do you see more 111 00:06:17.040 --> 00:06:19.759 people figuring it out and just architecting around it. 112 00:06:20.600 --> 00:06:25.399 Yeah, good question. So on for that that's really our 113 00:06:25.519 --> 00:06:26.279 sweet spot. 114 00:06:26.399 --> 00:06:28.959 I know it's counterintuitive because it's like, Okay, well here's 115 00:06:29.000 --> 00:06:32.759 this you know, startup. But the brands that we work 116 00:06:32.839 --> 00:06:36.439 with are like, you know, probably drove you to work today. 117 00:06:36.439 --> 00:06:37.959 I guess if you work from home, maybe not, but 118 00:06:38.079 --> 00:06:40.720 for those of you that commute to work, like you know, 119 00:06:40.759 --> 00:06:43.240 whether it's like the largest electric car company, the largest 120 00:06:43.279 --> 00:06:46.079 city and largest bank in the US, largest like iced 121 00:06:46.120 --> 00:06:49.560 B in Europe, one of the largest banks in Australia, whatever. 122 00:06:49.839 --> 00:06:51.480 When you look at the brands that we tend to 123 00:06:51.519 --> 00:06:55.000 work with there, they tend to be this this fortune 124 00:06:55.000 --> 00:06:57.560 two thousand and I think fundamentally what happened is twofold 125 00:06:57.639 --> 00:07:01.839 one AI has definitely accelerated the pace at which people 126 00:07:02.319 --> 00:07:06.079 are wanting to move towards a more interactive architectures. That's 127 00:07:06.079 --> 00:07:09.680 why too people realize that really the log is the 128 00:07:09.720 --> 00:07:13.000 source of truth in an architecture and databases are caches. 129 00:07:13.040 --> 00:07:15.480 And I think this wasn't my idea was it was 130 00:07:15.720 --> 00:07:18.959 popularized by Amazon They're like the log is the source 131 00:07:19.000 --> 00:07:21.439 of truth and everything else is just the cash and 132 00:07:21.480 --> 00:07:25.519 so uh and I guess you know, third where when 133 00:07:25.560 --> 00:07:28.519 we started talking about Iceberg and so on, people are like, 134 00:07:28.639 --> 00:07:33.160 finally I get finally all the blocks. Uh you know, 135 00:07:33.279 --> 00:07:35.120 queny when you put the final piece on your lego, 136 00:07:35.199 --> 00:07:38.480 it's just like it looks beautiful and it works. Is 137 00:07:38.480 --> 00:07:43.959 when you can start to glue analytics, which is the 138 00:07:43.480 --> 00:07:46.519 look looking behind the present, which is really what Redpana 139 00:07:46.680 --> 00:07:49.319 is with with Apachi Iceberg as the glue, and then 140 00:07:49.360 --> 00:07:55.000 the future with autonomous decision making with with agents by 141 00:07:55.040 --> 00:07:57.720 carrying context. I was like, now I think people understand 142 00:07:57.720 --> 00:08:00.800 the timeline horizon of how an application works, and so 143 00:08:01.720 --> 00:08:05.399 we're seeing the opposite. We're seeing actually acceleration in the 144 00:08:05.439 --> 00:08:11.519 fortune two thousands or five thousand really more more then 145 00:08:11.759 --> 00:08:13.560 I sell early on. And I don't know if it's 146 00:08:13.600 --> 00:08:15.839 a function of brand. Maybe we have better marketing, Like 147 00:08:15.959 --> 00:08:17.560 it's kind of hard to tell, and I don't want 148 00:08:17.560 --> 00:08:24.720 to like a you know, correlate and or incorrect assumptions here. 149 00:08:24.759 --> 00:08:27.399 And so I think fundamentally where if you were to 150 00:08:27.439 --> 00:08:29.839 ask me, what's like your gut, I think is the 151 00:08:30.040 --> 00:08:35.600 world is thinking about applications differently in large because of 152 00:08:35.639 --> 00:08:41.320 AI and in that world, red Panda place a critical role. 153 00:08:41.480 --> 00:08:44.480 And I'm not sure, you know, in the context of 154 00:08:44.519 --> 00:08:46.600 our by C for example, there there's like a. 155 00:08:48.200 --> 00:08:49.639 You know, many alternatives. 156 00:08:50.559 --> 00:08:52.600 You know, you just reminded me of something. And I 157 00:08:52.639 --> 00:08:54.799 haven't read too much about this. And We've talked about 158 00:08:54.799 --> 00:08:57.480 streaming for years. I've talked about streaming for many, many years. 159 00:08:57.600 --> 00:08:58.759 There's MQ series. 160 00:08:59.080 --> 00:09:02.159 There's all kinds of things in the past that were streaming, right, 161 00:09:02.200 --> 00:09:05.559 It's like little message cues basically MQ series all kinds 162 00:09:05.559 --> 00:09:08.240 of different message cues, which is kind of what's streaming. 163 00:09:08.840 --> 00:09:10.480 Is that that's the category that's in. 164 00:09:11.039 --> 00:09:14.000 But to your point, there is a maturation now in 165 00:09:14.080 --> 00:09:17.200 the environment, and I'm thinking to myself, this is a 166 00:09:17.240 --> 00:09:21.039 real time application architecture, right. That's what it really boils 167 00:09:21.080 --> 00:09:23.960 down to is like, guys, look, you want an application 168 00:09:23.960 --> 00:09:26.519 that does something. What is the fuel that's going to 169 00:09:26.600 --> 00:09:28.320 use to do that something. Is it going to be 170 00:09:28.440 --> 00:09:31.440 a traditional static database which is the traditional way of 171 00:09:31.480 --> 00:09:34.360 doing things, in which case, guess what, that database will 172 00:09:34.399 --> 00:09:36.559 slow you down, Like sooner or later, it's going to 173 00:09:36.559 --> 00:09:40.080 slow you down either because it's going slowly, or things 174 00:09:40.080 --> 00:09:42.759 that attached to it are going slowly. But the point 175 00:09:42.799 --> 00:09:47.399 is that this just ever ever growing ballast. If you 176 00:09:47.600 --> 00:09:49.200 kind of look at it in a certain way, do 177 00:09:49.240 --> 00:09:50.879 you really want that? And I think that's your answer 178 00:09:50.879 --> 00:09:53.919 is no, not really. What you want is for your 179 00:09:53.919 --> 00:09:56.879 apps to use real time data that is live right now, 180 00:09:56.919 --> 00:09:59.960 that is interactive with the users or the partners or whatever. 181 00:10:00.559 --> 00:10:03.360 And that's what you're enabling, right It's a real time 182 00:10:03.440 --> 00:10:04.799 application architecture. 183 00:10:05.919 --> 00:10:09.159 Yeah, and so I think two comments. 184 00:10:09.240 --> 00:10:12.279 The first comment is that I think the reason why 185 00:10:12.480 --> 00:10:15.799 database is not enough is because you lose data. And 186 00:10:16.080 --> 00:10:22.039 so set another way, you need the highest fidelity of 187 00:10:22.120 --> 00:10:25.600 the data to be able to reconstruct a different materialization 188 00:10:25.679 --> 00:10:29.080 of your data. And so if your data it's in 189 00:10:29.080 --> 00:10:31.399 a relational database, let me give you the example is 190 00:10:33.120 --> 00:10:37.440 average sales price and so or like the price of 191 00:10:37.480 --> 00:10:40.000 the last ticker symbol and ice that that's probably a 192 00:10:40.039 --> 00:10:42.639 better example. The amount of data that is flown with 193 00:10:42.720 --> 00:10:46.679 every tick of that symbol as it's being traded is tremendous. 194 00:10:46.879 --> 00:10:48.440 But when you ask a database, it gives you one 195 00:10:48.480 --> 00:10:51.080 killer by the data vice like this is the company name, 196 00:10:51.240 --> 00:10:53.480 this is the ticker symbol whatever. This is some price, 197 00:10:53.919 --> 00:10:56.159 but the amount of input to the database could be 198 00:10:56.159 --> 00:11:00.120 a gigaby per second. And so what happens is you 199 00:11:00.159 --> 00:11:02.639 lose data, right like fundamentally it is it is a 200 00:11:02.759 --> 00:11:06.120 lossy compression algorithm almost right, like you're sending your data, 201 00:11:06.159 --> 00:11:10.000 but you're only querying the last point in time. And 202 00:11:10.039 --> 00:11:12.480 so that's what the databases are. And so I think 203 00:11:12.840 --> 00:11:17.440 it is limiting less from like particular you know, I 204 00:11:17.440 --> 00:11:20.759 think through put databases have gotten really good at throughput costly, 205 00:11:21.840 --> 00:11:24.039 you know, so maybe the other limit is dollars, but 206 00:11:24.080 --> 00:11:26.960 they I think that now modern databases are pretty good 207 00:11:26.960 --> 00:11:30.279 at handling the throughput latency not not so good. And 208 00:11:30.320 --> 00:11:33.120 so but it's more fundamentally it's about like you're losing 209 00:11:33.279 --> 00:11:35.759 access to the real data. And so what happens is 210 00:11:36.519 --> 00:11:42.200 it forces the architect and engineer to have like, uh, 211 00:11:42.480 --> 00:11:44.879 computed all of the ways in which they wanted their 212 00:11:44.960 --> 00:11:47.279 data before they write an application. And what the log 213 00:11:47.399 --> 00:11:50.000 gives you is freedom. You're like, well, I want to 214 00:11:50.000 --> 00:11:52.559 store it in a really cheap storage. They's say S 215 00:11:52.600 --> 00:11:54.639 three or something like that compatible and then today is 216 00:11:54.679 --> 00:11:57.679 stick pulling, then tomorrow could be whatever. Elastic could be 217 00:11:57.720 --> 00:12:00.960 whatever it is, and so that freedom is just key 218 00:12:01.159 --> 00:12:04.840 I think. For that's where I see like the biggest 219 00:12:04.960 --> 00:12:07.639 pillar from an architectural perspective. 220 00:12:07.679 --> 00:12:08.759 That's very interesting. 221 00:12:10.639 --> 00:12:13.080 And then for AI thought that changes a little bit too. 222 00:12:13.159 --> 00:12:14.399 But that's about way for a second. 223 00:12:14.679 --> 00:12:16.080 Yeah, what then what do you mean by that? 224 00:12:16.960 --> 00:12:21.279 So Okay, Historically, I think we talked about the challenges 225 00:12:21.320 --> 00:12:27.639 of picking a database. Yes, in old school databases, perhaps scale, 226 00:12:28.039 --> 00:12:30.720 uh you know, on throughput and latency those for limitation. 227 00:12:30.919 --> 00:12:35.000 In modern cloud databases, I think throughput is okay. Latency 228 00:12:35.039 --> 00:12:38.000 is still not great and so maybe for some use 229 00:12:38.039 --> 00:12:39.960 cases does tend it to be. But we talked about 230 00:12:40.000 --> 00:12:42.440 like the architectural primitives that you get when you adopt 231 00:12:42.440 --> 00:12:43.720 the law as the source of truth. 232 00:12:44.720 --> 00:12:46.240 What happens is when. 233 00:12:46.679 --> 00:12:49.080 I'm going to define an agent as like an object 234 00:12:49.799 --> 00:12:52.279 with an L l M right, and so for simplicity 235 00:12:53.120 --> 00:12:58.600 and when you have these things, uh, what I guess 236 00:12:58.720 --> 00:13:02.240 the future is going to be about this autonomous decision making. 237 00:13:02.519 --> 00:13:06.120 And so if you look at a timeline t and 238 00:13:06.200 --> 00:13:09.159 you say analytics stuff like data breaks and so on. 239 00:13:09.320 --> 00:13:12.200 They were all about looking at the past, help me 240 00:13:12.320 --> 00:13:16.080 do my job better. Where during best sellers, what region 241 00:13:16.320 --> 00:13:19.399 is the best producing region, what product is the best 242 00:13:19.399 --> 00:13:21.600 producing product? What should be my average sales price? 243 00:13:21.679 --> 00:13:21.840 Right? 244 00:13:21.879 --> 00:13:24.919 Like, help me do what I do better by looking 245 00:13:24.960 --> 00:13:29.679 at the past. That's analytics. Red Panda has been classically 246 00:13:29.919 --> 00:13:32.879 and we'll talk about NX about like being the best 247 00:13:32.879 --> 00:13:35.200 operational system in the world. Right, So if you're running 248 00:13:35.480 --> 00:13:37.480 your system and you depend let's say you're trying to 249 00:13:37.519 --> 00:13:41.240 protect white house that go from getting taken down by 250 00:13:41.600 --> 00:13:44.879 North Korea bots or something like that, you need to 251 00:13:45.039 --> 00:13:48.399 know the movement of IPAs around the world, or maybe 252 00:13:48.519 --> 00:13:51.519 you're trying to track fraud detection or whatever. So that's 253 00:13:51.519 --> 00:13:55.519 the present. That's the operational sense. The future is about 254 00:13:55.519 --> 00:13:59.279 carrying the context of the past and presents so that 255 00:13:59.399 --> 00:14:03.320 this code can make autonomous decision for you. And so 256 00:14:03.559 --> 00:14:06.039 let me let me can now loop all the timelines past, 257 00:14:06.039 --> 00:14:09.440 present and future. And so the past and present are 258 00:14:09.519 --> 00:14:14.000 going to be bridged by technologies like apati Iceberg, which 259 00:14:14.039 --> 00:14:17.600 is a way to have a zero shot integrations with 260 00:14:17.639 --> 00:14:20.519 all the query systems, and so that that's an important 261 00:14:20.519 --> 00:14:23.960 tig piece and streaming engines will play a second. So 262 00:14:23.960 --> 00:14:26.480 that's where streaming engines play super critical role. Is just 263 00:14:27.120 --> 00:14:31.720 exposing the data in a way that is accessible for 264 00:14:31.840 --> 00:14:35.679 many many tools, right doctor B ClickHouse, right like red 265 00:14:35.720 --> 00:14:39.480 pand or whatever Pinot trino U. Basically just about every 266 00:14:39.519 --> 00:14:41.840 data bas is going to speak Iceberg. So that's that's 267 00:14:41.840 --> 00:14:45.279 a critical road for streaming engines. And then for the future, 268 00:14:45.320 --> 00:14:49.879 it's about carrying context for the agents to do their job. 269 00:14:50.039 --> 00:14:51.919 And so let's say that the job of an agent 270 00:14:52.000 --> 00:14:55.799 is to determine yes or no to a credit card 271 00:14:55.799 --> 00:14:58.480 transaction or a credit or a crediting period or whatever 272 00:14:58.919 --> 00:15:01.559 you need to give it cont and so think of 273 00:15:01.639 --> 00:15:03.399 a self driving car, right, Like you need to carry 274 00:15:03.399 --> 00:15:06.840 the context, either like the images or the previous five 275 00:15:06.919 --> 00:15:09.639 transactions or some sort of windows so that the agent 276 00:15:09.759 --> 00:15:11.879 can make a decision. So as part of the prompt, 277 00:15:12.279 --> 00:15:14.919 you can tell the agent, Hey, is this or is 278 00:15:14.960 --> 00:15:17.240 this not a credit card transaction? 279 00:15:18.440 --> 00:15:20.759 These are Alex's last five. 280 00:15:20.559 --> 00:15:23.480 Credit card transactions and he just bought a bagel in 281 00:15:23.519 --> 00:15:26.799 San Francisco. So if it gets transacted in the UK, like, 282 00:15:26.840 --> 00:15:29.480 obviously market has fraud, and give me the explanation so 283 00:15:29.519 --> 00:15:31.639 I can render it on the screen. Right, And so 284 00:15:31.879 --> 00:15:35.840 I think that that's why streaming engines are seeing in acceleration, 285 00:15:36.039 --> 00:15:41.799 one as a glue to analytics and the present with Iceberg, 286 00:15:41.879 --> 00:15:44.200 and then second as a way to carry context for 287 00:15:44.240 --> 00:15:45.679 this autonomous decision making. 288 00:15:46.320 --> 00:15:49.919 So I'm trying to processes my brain, right. So when 289 00:15:49.919 --> 00:15:52.519 you have, like with cough gets topics, right, you have 290 00:15:52.600 --> 00:15:55.000 topics that are streaming and you can choose which ones 291 00:15:55.039 --> 00:15:58.360 you pull together in any given point in time as 292 00:15:58.399 --> 00:16:01.120 I recall, and that be comes the fabric or the 293 00:16:01.120 --> 00:16:02.679 context that you're talking about. 294 00:16:02.919 --> 00:16:04.200 So here you're. 295 00:16:04.080 --> 00:16:08.080 Talking about getting context from log files basically from systems 296 00:16:08.480 --> 00:16:12.159 that are important to this particular potential transaction. And what 297 00:16:12.200 --> 00:16:15.519 you're talking about is allowing the agent to be a 298 00:16:15.639 --> 00:16:22.120 much more adaptive and real time semi autonomous entity that 299 00:16:22.200 --> 00:16:25.399 will absorb context and just make a decision quickly. Yeah, 300 00:16:25.559 --> 00:16:29.200 Like that's the idea, is that? So you've in a sense, 301 00:16:29.279 --> 00:16:35.639 you've deconstructed the traditional data flow that goes into an 302 00:16:35.679 --> 00:16:38.240 application which makes a decision. Because to your point, like 303 00:16:38.480 --> 00:16:40.440 all the Stata warehousing, that's my background, it's all the 304 00:16:40.440 --> 00:16:43.559 Stata warehousing stuff, right, you do ELT and et and 305 00:16:43.559 --> 00:16:44.000 all this stuff. 306 00:16:44.000 --> 00:16:44.720 You get into this. 307 00:16:44.960 --> 00:16:48.799 Big central repository, which we did because they figured out 308 00:16:48.799 --> 00:16:51.279 that you couldn't really query an ERP It's not what 309 00:16:51.320 --> 00:16:53.879 it was designed for. And besides, what you want to 310 00:16:53.879 --> 00:16:55.919 do is know how all this stuff relates to each other. 311 00:16:56.000 --> 00:16:58.000 That's where the magic is, right, is how all these 312 00:16:58.039 --> 00:17:01.519 different pieces of data relate. And so with red Panda, 313 00:17:01.879 --> 00:17:05.240 what you've done is expedited the process of feeding the 314 00:17:05.279 --> 00:17:09.559 important bits from important systems into a real time context, 315 00:17:09.920 --> 00:17:12.559 which the agents can then grasp as needed. 316 00:17:12.599 --> 00:17:13.119 Is that right? 317 00:17:13.759 --> 00:17:17.200 So it's cool about red Panda From an engineer perspective, 318 00:17:17.200 --> 00:17:20.279 it's basically most of our code is in the open, uh. 319 00:17:20.319 --> 00:17:24.000 And so we will link an example here of what 320 00:17:24.039 --> 00:17:27.759 I mean, like a simple concise yamal file so people 321 00:17:27.799 --> 00:17:29.720 can just like see the whole thing put together and 322 00:17:30.039 --> 00:17:33.000 it becomes like really becomes something that they can, like 323 00:17:33.039 --> 00:17:34.119 they can run on their laptop. 324 00:17:34.160 --> 00:17:36.880 But that's exactly right. And so. 325 00:17:38.519 --> 00:17:40.079 First of all, I think it's worth it just to 326 00:17:40.119 --> 00:17:42.119 share like maybe a little bit of redpan So we 327 00:17:42.200 --> 00:17:46.079 did start red pandas is like you know, high performance 328 00:17:46.400 --> 00:17:49.759 storage engine. And the idea is like if we adapt 329 00:17:49.799 --> 00:17:50.480 to how. 330 00:17:50.559 --> 00:17:52.319 The hardware is. Right, so. 331 00:17:53.880 --> 00:17:56.759 Most most software just operates on like you know, a 332 00:17:56.799 --> 00:18:00.880 tremendous amounts of layers of abstraction, and you'll loose performance 333 00:18:01.000 --> 00:18:03.960 with every layer because you're generalizing things. And so when 334 00:18:04.000 --> 00:18:07.599 I started a red panel's like, what if we just 335 00:18:08.680 --> 00:18:11.200 adapt to how hardware works, and if you go all 336 00:18:11.240 --> 00:18:14.839 the way down to even like core to core coherency protocols, 337 00:18:14.839 --> 00:18:17.680 so you have you know core they say L zero 338 00:18:18.000 --> 00:18:21.319 or you know zero to one on a dual core system, 339 00:18:21.880 --> 00:18:24.039 single socket, right, so single motherboard. We don't have to 340 00:18:24.039 --> 00:18:26.599 complicate the picture with multiple sockets on a motherboard, but 341 00:18:26.720 --> 00:18:30.480 one motherboard you have one on one chip and the 342 00:18:30.559 --> 00:18:33.440 chip has two dual core right when you go down there, 343 00:18:33.440 --> 00:18:37.119 when you're talking about memory coherency protocol, it's all message passing. 344 00:18:37.680 --> 00:18:41.400 And the insight here from hardware manufacturing people is that 345 00:18:42.279 --> 00:18:44.200 you could do a bunch of useful work if you 346 00:18:44.240 --> 00:18:48.240 don't block, right, And so if you adopt a similar 347 00:18:48.640 --> 00:18:53.440 software architecture where it's all message passing, you can yield 348 00:18:53.599 --> 00:18:57.240 at points of blocking and then you can let the 349 00:18:57.279 --> 00:18:59.039 CPU do a lot more useful work. 350 00:18:59.079 --> 00:19:01.799 And so that's really interesting. Folks, don't touch it. That 351 00:19:01.799 --> 00:19:02.519 will be right back. 352 00:19:02.640 --> 00:19:09.759 You were listening to inside Analysis. 353 00:19:11.759 --> 00:19:19.799 Welcome back to inside Analysis. Here's your host, Eric tabanac. 354 00:19:19.400 --> 00:19:23.640 Red pand that runs on a network of ny square 355 00:19:24.039 --> 00:19:27.880 single producer, single consumer, lock free cues that acts as 356 00:19:27.920 --> 00:19:30.599 mailboxes for every core. And you can assume that every 357 00:19:30.599 --> 00:19:34.480 core is like an actor system on a super relatency network. 358 00:19:34.920 --> 00:19:37.480 And so when you embrace that, by the way, you 359 00:19:37.519 --> 00:19:43.839 get a simple concurrency I guess software structure and a 360 00:19:44.000 --> 00:19:50.960 simple parallelism uh deployment and so anyways, performance and this 361 00:19:51.160 --> 00:19:53.119 is like one of my favorite topics to talk about. 362 00:19:53.240 --> 00:19:54.960 I want to stop us here because this could be 363 00:19:55.000 --> 00:19:58.160 a two hour call where we like dig super deep 364 00:19:58.160 --> 00:19:59.920 into the mechanics of. 365 00:19:59.880 --> 00:20:03.119 The This is so interesting for lots of different ways. 366 00:20:03.160 --> 00:20:04.799 I mean, I've been around this business a lot of time, 367 00:20:04.839 --> 00:20:06.680 so I learned from all sorts of different vendors about 368 00:20:06.720 --> 00:20:09.640 different things they're doing, and it's always very interesting. But 369 00:20:09.720 --> 00:20:13.079 in the recent past, meaning six seven months ago, I 370 00:20:13.160 --> 00:20:14.680 ran into a company called hammer Space. 371 00:20:14.720 --> 00:20:17.000 Are you familiar with hammer Space that I mentioned those guys. 372 00:20:17.799 --> 00:20:19.559 You need to look into these guys, I mean, because 373 00:20:19.599 --> 00:20:22.240 they are thinking in a very similar way that you're thinking. 374 00:20:22.759 --> 00:20:25.359 They went out what they do. They're a parallel file system. 375 00:20:25.680 --> 00:20:28.119 So this is big in the HPC world for life 376 00:20:28.160 --> 00:20:31.279 sciences and training models. They're actually training a LAMA two 377 00:20:31.319 --> 00:20:33.839 and LAMA three, so they're being used to deliver the 378 00:20:33.920 --> 00:20:36.039 data to the CPU, and they did something that was 379 00:20:36.039 --> 00:20:38.759 so damn clever it blew my mind. He talked about 380 00:20:38.759 --> 00:20:43.559 how in the average AI training architecture, data has to 381 00:20:43.559 --> 00:20:46.200 make eleven hops to go from where it is to 382 00:20:46.279 --> 00:20:49.960 the GPU. It's like controller nodes and different things like 383 00:20:50.000 --> 00:20:52.319 that in the storage environment, the network, all these different 384 00:20:52.319 --> 00:20:55.440 steps along the way, but still that's eleven freaking hops. 385 00:20:55.640 --> 00:20:56.440 So they talked. 386 00:20:56.200 --> 00:20:59.279 About how I think it's vast to cut off one 387 00:20:59.400 --> 00:21:01.960 hop and cumulo cut off two hops. 388 00:21:02.000 --> 00:21:03.960 Maybe, but these guys. 389 00:21:03.640 --> 00:21:07.880 Were able to figure out by affinitizing the processing to 390 00:21:07.960 --> 00:21:12.119 where the GPU server, which has built in flash memory 391 00:21:12.480 --> 00:21:16.160 that often gets ignored. They call that stranded capacity basically 392 00:21:16.200 --> 00:21:19.640 because a lot of times companies will want to just 393 00:21:19.799 --> 00:21:22.720 have all their governance protocols and security around the storage 394 00:21:22.799 --> 00:21:24.640 array and they don't want to have to try to 395 00:21:24.680 --> 00:21:27.200 figure that out on the GPU array as well, right, 396 00:21:27.240 --> 00:21:29.920 so it's these two separate worlds. Well, they figured out 397 00:21:29.920 --> 00:21:34.119 how to affinitize that to where when the processor knows 398 00:21:34.240 --> 00:21:36.359 that the data is on this machine, Well, I don't 399 00:21:36.400 --> 00:21:37.480 even have to go to the network. I don't have 400 00:21:37.519 --> 00:21:39.319 to go to a storage around Look anywhere is right here. 401 00:21:39.880 --> 00:21:43.359