00:01:42 · Rajkumar Sen
[00:01:42] S1: So what happened is that you know you have to go back like uh to like the days when Databricks and Snowflake were early, right? Yeah. So when AWS came, like um, I know the likes of Oracle and like you know, cloud is like, you know, it's not a non-starter. But then people realized that, you know, uh it gave people the opportunity to build scalable, you know, applications. And then I think every aspect of the data data space started to like change. So new databases, for example, right? There are two streams that came along. One is the distributed stream where Cockroach DB, Yugabyte, uh, and maybe a couple of other companies. Uh, and then the and then as uh S3 became more and more like you know um reliable, uh there was a push to separate compute and storage at the warehousing level, which Snowflake and Databricks did. And at the database level, Neon obviously was a pioneer in that, and now I think others are also doing things. So the trace of all of this, the origin of all of this is like the innovation done by AWS at some point. If S3 were not that powerful, none of this was materialized. Then things around the database also need to evolve. Like the how do you move data from A to B? Do you move in files or do you move events, right? So Arcion, uh, when we started, we called the company was named as Blitz because we are like, you know, data has to travel super fast between system A to S to enable real-time machine learning, real-time BI. You know, it was like a technical insight. Uh, then how do you store that kind of data, event data from databases and others? You know, Kafka was getting popular. Uh and so I think you need like that fundamental change in hardware and software and the compute to be able to, or in very new way to, you know, um uh process things. So the cloud was that you know, event that uh databases got changed. Things around databases like data movement, ETL got changed. Um event, you know, um, the way to sort events was IBM MQ and you know, uh Tibco and all of that. Uh they need to be more scalable, confident, build a great solution around it. Uh, and now we will see the next set of waves because now, alongs with all of that, now you have you know agent applications and um non-humans called agents wanting to write software, generative new software on the fly. So there will be a whole overhaul of the infrastructure space.
00:04:21 · Rajkumar Sen, Siddhartha Ahluwalia
NEON is very interesting because uh when Neon started, uh, it the aim of Neon was to uh enable AI users to uh you know spin up new databases and you know do branching and all of it. And uh fortunately for me, I used to work at a database startup called MemSQL, which was Nikita Shamgunov's first database company. So I've been following Neon from day one. Uh Nikita and I both were common investors at Mooncake, also, which you know uh got tackled with Databricks. So I was a big supporter of Neon. Uh, we were actually partnered also. Sure. Uh so but I think there'll be more. I think Databricks is innovating a lot. They've released this new LTAP thing recently. Uh Monday, Clickhouse announced a big thing. Uh Andy Pavlo, arguably, you know, one of the top professors of data business at Carnegie Mellon. He joined Clickhouse as the vice president. [00:05:18] S0: One of my friends' company, PeerDB, that were acquired by Clickhouse. [00:05:23] S1: Yeah. [00:05:24] S0: And I think those people who stuck to Clickhouse when they acquired when Clickhouse was one billion.
00:05:29 · Rajkumar Sen
[00:05:29] S1: Yeah, yeah. No, I know Sai. Yeah. Uh I think he lives in San Carlos. Uh so we are connected. Uh, I'm super happy. Uh, though also, you know, we would have been happier if he would have, you know, he would have come and joined us. But uh and uh click house, uh Alexei actually I know for a long, long time when he was at Yandex. Okay, because I also was also doing database research for a long time. So I just spoke with him yesterday after they announced the Andy Pavlo thing. So I think there's going to be more innovation around it because a lot of database features have to be AI like assistant, and also really AI features that have to be, you know, maybe somebody is going to push SLMs inside the database, you know, instead of taking the data out and that compute inside Postgres or something, somebody will do it. Databricks now flick click house. So there's going to be that, you know, uh uh, and or maybe super basic. They're also trying to build this uh next generation Postgres. So there will be this uh there is going to be a lot of data management innovation. I also feel like uh the whole uh conference space also may get disrupted. Uh we don't know when. Yeah. This is very hard to predict. Um anytime there is like a new computing paradigm or hardware and software change, uh, there is always a little bit of money and to be made in the data management space, which is why I like this space. You know, um databases are not constant. If that way true, everybody was used Oracle, yeah, and then everybody was used MongoDB. Yeah. And now everybody shifted to Postgres. So we're seeing three waves, you know, first the Oracle IBM wave. I used to work at Cyclist SAP, then the unstructured wave, MongoDB, because Santa, everybody thought everybody's gonna move to unstructured, then came the Postgres wave. So we're gonna see a fourth wave. I wish I could predict what the fourth wave is, but there is going to be a fourth wave. Uh, maybe a single data management layer that can do event and unstructured for AI. Uh, but there are a couple of things I would highlight, since we're still topic, right? I think um as I'm looking into this, and I think I spoke about it also with a few people at the Databricks and Snowflake Summit, that there is one thing I'm seeing which will require databases to significantly overhaul their capabilities.
00:07:54 · Rajkumar Sen
See when if you are a human and you're writing a query, you are writing a query, or you know, your PI tool is firing the query, the query is running slow, yeah, or the query is giving wrong results. I think there is still a defined process, human-driven, to figure out what's going on. Okay, why my results are wrong? I feel the results are wrong. Talk to another person who knows the data. Uh, if an agent is doing that, like if the results are not returning in five minutes, the human probably has the patience to deal with it. Will the agent be trained to deal with it? Like, because the agent's probably uh waiting to curate that, join it as somewhat, and give the response to a question, right? So it's very different now that uh I think uh the response of a database to any query, yeah, search query, whatever, has to be super low latency because a lot of those queries will be you know joined together to give the final answer to that. So there's like a post-query distillation and joining aggregation happening. Uh now data availability has to be real time, your results have to be real time, have to be accurate. I mean, I'm some companies have started to, you know, they're saying, you know, for Text to SQL, we we have to give accurate results, no-brainer. But uh all of that has to also happen at the database level. Like you know, you're finally running the query somewhere on top of iceberg or somewhere. So, how do you give answers in milliseconds for 20 queries and then the agent is low? So a lot of innovation has to happen. They have to be fast, they have to be reliable, and they have to be repeatably and reliable. Like you can't have like the first thing answered in one second, and second one is taking five seconds. Both of them have to give exactly the same answer. So I think uh we will see a lot of innovation, and most of the innovation will be driven by applications forcing this thing, um, which explains why people have spent one billion on Neon and 250 million on Crunchy. And a lot of innovation is also being done at uh Microsoft on Horizon DB, also. Uh, one of my ex uh bosses uh from MemSQL could actually make it as first company. He's leading the Postgres sector. So, I mean, as much as I'm not in deep data business, I keep following like what's there.
00:10:14 · Rajkumar Sen, Siddhartha Ahluwalia
Uh but um I'm excited to see you know uh new companies also doing a lot of things like Lake House and others. [00:10:21] S0: So, besides this speed, what was the differentiated insight in Arcion that you started with and that gave you the initial traction?
00:10:30 · Rajkumar Sen
[00:10:30] S1: See when you start a company as a first-time founder and as a solo founder, so you know the journey, right? How hard it is. So when we started, we said, you know, if you have a use case that others are not doing, Oracle is not solving, or AW is not solving, we'll solve it. It has to form within the structure of real-time data delivery. Yeah. Now the use case could be an OLTP system to an OLAP system. Now, if you think of 2018-2019, there was not a single OLAP system which was consuming events. All of them were very happy. Hey, give me a CSV file or a parquet file. Actually, if they could ingest parquet file properly in 2018, it was considered to be a big achievement. CSV was a de facto. I will not name which data warehouses are popular in 2019, 2020. You know the answer. All of them were saying we are happy with files. So, what do we do with a modern technology like ICU, right? Where we are saying, so we are looking out for use cases where event-driven transactional events meant business outcomes. Yeah. Fortunately, um, you know, one of the largest cybersecurity companies in the Bay Area, which is public now, uh, I don't know whether I can share. Uh, we met them at a conference. They were running one of the largest MongoDB clusters in the whole world. It was a uh they're cutting all the security events for all their customers. Think of like a cybersecurity company who is deployed at DoorDash, Meta, everywhere, collecting real-time security events in a uh MongoDB cluster, like 400, like huge cluster. Yeah, but what but they said hey, you know, we have a problem because when we have a breach, then we have to create a copy of that tenant for forensics, yeah. But we still have to keep collecting so that required them to also have a uh real-time replica of their MongoDB cluster with two or three seconds of latency. Yeah, and you know the PRDB use case also we're being investors similar. This is 2018-2019 time frame, and they said um when the incident happens, we want to quarantine and give snap, you know, the replica is available to the customer to do forensics. Interesting project. Of course, it's the Bay Area company with super smart engineers. Uh, MongoDB folks said we are working on a project that will give you this thing. And these set of solutions existed in Oracle and all before, right? With GoldenGate and others.