James Morle

Posted without comment

@gwenshap Ghosting them and invalidating their query results?

What's your preferred way of isolating databases?

@MarcJBrooker I’ve been using it for two days, and I’m super impressed!

Check out Kiro! Kiro moves the AI IDE from 'vibe coding' to development driven by specifications. Specs allow AI-powered development that's repeatable, scalable, and deals better with large and complex projects.

I’m looking for ‘interesting’ silicon wafers for my personal mini-museum. CPU, FPU, GPU, *ROM, *RAM, historic and new. Please DM or reply if you can help! Any RTs highly appreciated, to help spread the word!

@MarcJBrooker Perhaps the intent here is that the MCP APIs are suboptimal for serious database access? I can get behind that, they are quite inefficient and naive currently. But as to whether agents should access prod data - they absolutely should, that’s the whole point.

I understand where this is coming from, but it's missing how useful relational databases are as tools for agents. Durability, transactions, queries, schema, all of these things are useful tools for agents that need to write stuff down and find it later.

@davepl1968 @marklucovsky In my Facebook days, I would see people leaving the campus with a stack of 4-5 togo boxes of dinner. Feed the whole family, or all your friends, it’s ok! It’s hard, living on the breadline…

@marklucovsky And on Fridays, I'd see people filling backpacks and gym bags with soda on their way home for the weekend. People that I KNOW could afford their own soda. Which is why we can't have nice things, and why socialism never works.

That's the problem with free juice. Eventually, you run out of other people's juice.

Case-in-point: My Ubiquiti network management app is very nice, but it makes it really difficult to ask my LLM to just work out what's wrong with my AP config that causes interference.

Epiphany of the day (for me): All user interfaces now need a "just dump everything in text format and download" method of viewing settings. Why? Because LLMs can digest it easier that way.

DynamoDB and Aurora had a baby -> Aurora DSQL. This is what you need to know. For decades, scaling relational databases meant trade-offs: - Read replicas for reads. - Sharding for writes. - And a pile of glue code to fake global consistency. That’s the reality we live with. But Aurora DSQL promises pure scalability; no hacks, no compromises. Aurora DSQL is a new serverless SQL database, optimized for transaction processing, and designed for the cloud. It just went Generally Available, and it's rewriting how we think about scale in transactional systems. Built on PostgreSQL, powered by a distributed architecture, and backed by Rust for low-latency performance, Aurora DSQL brings active-active, multi-region SQL to the mainstream. All the SQL stuff you expect is there: transactions, schemas, indexes, joins, and so on, all with strong consistency and isolation. Just one endpoint per region. Write anywhere. Read anywhere. Strongly consistent. Why this matters: - Write scalability without sharding. - True multi-region failover without downtime. - PostgreSQL compatibility; your apps don’t need rewriting. - Auto-scaling without provisioning complexity. And yes, it feels like DynamoDB and Aurora had a baby, and it speaks SQL. We’re witnessing a shift where distributed SQL isn’t just research material or niche; it’s production-ready, cloud-native, and cost-aware. I think this unlocks a new design space for systems that need transactional guarantees and global scale, without having to choose one. Is this the end of database hacks in large-scale systems?

@apparentorder @astuyve I'd be happy to investigate this further! Can you ping me your cluster ID over DM, please?

According to Cloudwatch, my little experiment is clocking in at ~1.4k DPU per hour (so about 1M DPU / month), almost all of it from about 10 MB reads per minute, according to the "BytesRead" metric. I'll need to understand this better, because AFAICT it should only be reading ~2k tiny rows every minute from an index scan, so 10 MB/min seems off and could become expensive quickly.

NEW: Amazon Aurora DSQL is GA with new pricing announced:

Happy to share that Aurora DSQL is now Generally Available!

@gwenshap Some of the recent cardinality estimation research is very interesting. Machine learning brings a CBO similar levels of insight to a skilled developer that really understand the data patterns. Inference time is still a challenge though.

@JamesMorle Yeah. As you know, optimizers that mis-estimates the cardinality in complex queries is a long standing issue in many DBs. And Postgres does not have hints natively!

Doing joins in the app is usually an anti-pattern. Unless your database is very very busy and you need very consistent latency…

LLMs have been transformational to my workflow. I didn't realize how much until my API access was disrupted for a day last week. I missed the pure labor saving stuff the most: - Topping and tailing code with arguments and help messages - Coding ANYTHING that accesses databases

@connor_mc_d ORDER BY dbms_random.random();

Whatever your politics, the DBA in me is screaming out "Where the hell is the ORDER BY clause?!?!?!"

Other takeaway is how AWS seems to have played GenAI just right with Bedrock (even if they have no own high-quality LLMs) This team looked at options and settled on Bedrock. It's secure, they trust AWS, & doesn't train on your data. You can choose your model. I hear so many companies who are a bit more conservative/worried about their data and: 1. Would want to host their own LLMs... 2. ... but it's a lot of work and is expensive 3. Hear about Bedrock. "Oh, it's what we need" Home run by AWS

@MarcJBrooker Nobody has seriously run Oracle databases on top of a filesystem for decades, for lots of very good reasons, including the ones in this thread. Then you have other good reasons such as the cost and futility of buffering a buffer.

Which suggests that it may be time to just get kernels and filesystems out of the storage business entirely for these kinds of applications.

In my mind, there are two real take-aways from issues like this. First, abstractions like POSIX (and even Linux specifically) are making it harder for database to take advantage of the semantics of their storage devices. This is the opposite of what good abstractions do!

Euclidean might help for non-linear layouts, like Hs and Ys, but only if there is a way to walk the Euclidean path!

Airport gate numbering is missing a trick: for some terminal layouts, the number interval between gates could be proportional to distance. If Gate 1 has a gate adjacent to it, call it Gate 2. If it’s 10x further, call it Gate 10. If it’s 100x further call it Gate 100.

@Chris_Mellor “Woncshe more, we play our dangeroush game, a game of chessh…”

The hunt for ....