georgewfraser

Slack is such a simple product, and is so strategic as an interface for AI, Anthropic has to be considering building their own Slack. Hopefully this tag approach is an MVP and it proves the potential of workplace messaging as an interface for AI, but also the limitations of relying on the extension points they salesforce chooses to provide, and it turns into a full fledged slack competitor from anthropic soon.

I assure you I wrote it myself

Claude-in-Slack is a big enough feature to overcome the slack-connect network effect. Openness is absolutely key! I wrote this post because I hoped that if Anthropic is already planning to do this I might be able to influence them to make open-data part of the plan. But openness by itself isn't a big enough feature to get users.

Also true! The most important thing is that the NewSlacks commit to interoperability. I think Anthropic has a special opportunity to lead the way here, because they have a track record of standing by their principles to an extraordinary degree.

You can only access public channel data, you can't even access that at scale, and Claude needs to be more natively integrated in ways that Slack will never allow.

I talked to the timescale CTO at pg conf a few years ago and asked him what timescale does differently than a standard columnar database that makes it better suited for time oriented data. He said a bunch of things and I said “but columnar databases do those things.” Then he got mad at me.

I guess it’s just another columnar dbms after all?

They make a really good criticism of Iceberg: if we have a database anyway, why are we bothering to store metadata in files?

I don’t think DuckLake itself will succeed in getting adopted beyond DuckDB, but I would not be surprised if over time the catalog just absorbs the metadata, and the original Iceberg format fades into history as a transitional form.

This is exactly right. We even went so far as to build a proof of concept internally, and the technical challenges are just very different. The simplest way to explain it is that Fivetran connects a skinny pipe (APIs) to a fat pipe (databases) while Census connects a fat pipe to a skinny pipe.

I am generally a huge vertical sharding skeptic but there are special cases where it is beneficial. If you have a simple query pattern on one table that represents a big fraction of your entire workload you can put it into its own instance and it becomes much easier to monitor. It’s easy to see why vertical sharding is sometimes the right answer by inverting the decision: should we put two unrelated large applications on the same instance? Obviously not, there is no benefit and ops becomes more difficult.

Like so many things from Google engineering this will be toxic to your startup. SREs read stuff like this, they get main character syndrome and start redoing the technical designs of all the other teams, and not in a good way.

This phenomenon can occur in all “overlay” functions, for example the legal department will try to run the entire company if you don’t have a good leader who keeps the team in their lane.

In general, I absolutely agree with you. It’s basically an instance of “the customer is always right”: if a smart customer can’t get our product working, there is a problem with the product. But this post made a much bolder (and wrong) claim: “the product has a number of major design flaws that mean that it literally cannot work”.

I have some insight into this because this claim is about my company Fivetran:

“…relies on the data source being able to seek backwards on its changelog. But Postgres throws changelogs away once they're consumed, so the Postgres data source can't support this operation”

Dan’s understanding is incorrect, Postgres logical replication allows each consumer to maintain a bookmark in the WAL, and it will retain the WAL until you acknowledge receipt of a portion and advance the bookmark. Evidently, he tried our product briefly, had an issue or thought he had an issue, investigated the issue briefly and came to the conclusion that he understood the technology better than people who have spent years working on it.

Don’t get me wrong, it is absolutely possible for the experts to be wrong and one smart guy to be right. But at least part of what’s going on in this post is an arrogant guy who thinks he knows better than everyone, coming to snap conclusions that other people’s work is broken.

You can partition your BigQuery table however you like and Fivetran will leave it in place. I don’t think there’s any benefit to partitioning the staging table.

I wonder if it would be easier to create a C virtual machine that emulates all the OS interaction, then recompile Postgres and the extensions to run on this. Perhaps TruffleC would work?

https://dl.acm.org/doi/10.1145/2647508.2647528

I have an S with a yoke and prefer it to a round steering wheel. 95% of the time it’s better: you can see the entire dashboard and your body can feel the angle of the wheel instinctively. 5% of the time it’s worse, during low speed maneuvers.

The world would be a better place if database drivers were completely abandoned as a way for clients to connect to databases. A standard API, implemented by multiple vendors, is a vastly preferable solution. Arrow Flight is an example of this.

https://arrow.apache.org/blog/2019/10/13/introducing-arrow-f...

Licensing under AGPL will make it hard for any startup to use Cozo. Lawyers always ask about AGPL in venture financing diligence and it is considered a red flag. You can argue that they are wrong, the linking exception and so on, but you’re basically shouting into the wind.

The core claim of this article, that Snowflake doesn't implement optimizations that would reduce usage, is not true. Search optimized tables, partitioned tables, and per-second billing are all counterexamples.

Yeah I think that is the key question: will data lakes become the dominant paradigm? There is certainly a lot of talk around them, though I see a ton of companies are still just going all in on a conventional data warehouse, but they tend not to talk about it because it’s not a new or interesting thing to do.

The thing I wonder about with Presto and to a lesser extent Spark is, how many of their users adopted this tool because it was an easy migration path from Hive, and how many of those users will eventually re-platform to something else?

There is no better way to get shut out as a scientist than to disprove the career-making findings of the important people in your field. One of the side effects of this is the “discussion does not match the results” paper. It’s routine to read papers where the discussion blatantly contradicts the papers own results. What is happening is the authors are disclaiming their own findings in order to get past hostile reviewers, betting that astute readers will notice the contradiction, ignore the discussion and draw their own conclusions from the results.

I hope not. We need to hire a Postgres DBA at Fivetran, we have more or less a single Postgres database with all our state, and it’s become clear that we need a full time person to optimize it.

Going to hazard a guess that he’s mad that Tesla isn’t a customer of his embedded-software-testing product https://www.ghs.com/

This is such a huge problem. It's even worse than it looks: because users are slow to upgrade, changes to the database system take years to percolate down to the 99th percentile user. The decreases the incentive to do certain kinds of innovation. My opinion is that we need to fundamentally change how DBMS are engineered and deployed to support silent in-the-background minor version upgrades, and probably stop doing major version bumps that incorporate breaking changes.

It is amazing how many large-scale applications run on a single or a few large RDBMS. It seems like a bad idea at first: surely a single point of failure must be bad for availability and scalability? But it turns out you can achieve excellent availability using simple replication and failover, and you can get huge database instances from the cloud providers. You can basically serve the entire world with a single supercomputer running Postgres and a small army of stateless app servers talking to it.

Wolfram wolfram wolfram wolfram! Wolfram’s wolfram wolfram, wolfram wolfram. Wolfram wolfram? Wolfram wolfram wolfram wolfram.

I would love to run a time-series benchmark against a good column store like Snowflake to see if purpose-built time series databases are actually faster. I have a sneaking suspicion the time scale databases are just reinventing the column store, and that an appropriate non-sabotaged benchmark would show this.

I’ve bought two monitors that I really loved: an Apple Cinema Display 27” for $2000 in 2003, which was a crazy amount of money for a monitor at the time. I used it for 12 years. And an Apple Pro display with all the options for $7k last year, which seemed a little crazy but it’s just unbelievably good, like looking through a window, and I won’t be surprised if I use it for 10 years. If you’ve got the money, a really good monitor is worth a lot.

It’s at least $1,000/sqft all-in. Construction costs have gone up a lot in recent years, and architecturally designed houses like this are expensive to design and build.

It’s too bad they didn’t share their final costs, there’s very little public info on how much custom builds like this really cost. I think people are a little embarrassed by how much they spent and so they don’t want to reveal it.