Building the self-driving Postgres we always wanted

One of the areas we adjust with almost every singe customer is reconfiguration of log_xxx parameters – this is simple and needs to be done proactively in each startup, to have data for future RCAs defaults are not enough!

Making Postgres slower as a service is it a good idea? cloud.google.com/blog/product...

BYOx is the new term (invented by claude in cursor during postgres_ai copilot development) -- it means "bring your own anything" can be cloud (BYOC) can be your own machine (BYOM) can be nothing -- we'll do it for you (our cloud) this is the way

We went full in with monospace web What started as local redesign of PostgresAI Console's left menu in line with copilot workflow (Monitoring → Checkup/Issues → PR/MR → Monitoring), turned out to be full redesign of the whole Console. Rolled out to all 11k+ users And I like at a lot dark mode WIP

among all similar screenshots you can see today, this one is out of questions

quick check of pg18 pgbench on my new m5 macbook pro with 10 vcpus – not bad, 0.5M select-only TPS 50k in-mem SELECTs per core is actually good -- for example 4th-gen EPYCs showed only ~10k per core postgres.fm/episodes/fou...

switch to main and pull

Just sent first 100 invites to PostgresAI copilot Senior DBA on autopilot Cyber Monday deal still open Want in? DM me

"hit/read ratio" should be called "hit/(hit + read) ratio" so the math matches or hit/nullif(hit + read, 0) to be safe

Added a new diagram to the postgres.ai front page showing how PostgresAI copilot works: Your Postgres (located anywhere) → postgres_ai monitoring → Health check & Issues → GitHub PRs / AI tools → deployed back to PG (Inspired by turbopuffer's clean design; it rotates 90° on mobile.)

you don't need AI to tune checkpoint_completion_target just set it to 0.9

No PostgresFM podcast this week – was busy with trips and work, see you next week, and THANK YOU for following and listening!

Tools show green "CPU" when it's actually I/O. And COPY is what pg_dump and pg_restore use by default. We went in to add wait events for COPY TO/FROM PROGRAM, then realized the gap is much bigger than expected. Verified with pg_wait_sampling. 1/

Hacking Postgres online – join! www.youtube.com/watch?v=wLPD...

Asked AI to analyze what wait events are missing in current PostgreSQL code – after a few iterations, accumulated > 70 gaps, here is the report: github.com/NikolayS/pos... Checking with -hackers if it makes sense to dig further: www.postgresql.org/message-id/f...

I like how easy it is now to get the new postgres.ai copilot for any PostgreSQL: - One-click deployment in three modes – PaaS, BYOC, BYOM - Easy access to dashboards at xxxxxx.pgai.watch by default – to dive into all the details - "Resolve in Cursor" (or you can use Claude Code, Codex, etc.)

Today is the day -- end-of-life for Postgres 13 www.postgresql.org/support/vers...

postges_ai monitoring 0.12: SLRU, MultiXact member space exhaustion, many improvements and fixes GitHub: github.com/postgres-ai/... GitLab: gitlab.com/postgres-ai/...

#PostgresMarathon 2-013: Why keep your index set lean postgres.ai/blog/2025111...

See it all the time: developers want faster DBs but only add indexes – never drop unused or rebuild degraded ones. This is why postgres.ai is powerful: it detects indexes that need to be dropped, sets up automated reindex, and observes the results. So you can work on features, not DB maintenance.

#PostgresMarathon 2-012: Ultra-fast replica creation with pgBackRest postgres.ai/blog/2025110... Couldn't believe the numbers: 35 TiB/h copying PGDATA from one EC2 instance to another. That's fast enough to clone a 10 TiB production database in ~17 minutes. 1/

Redesigning @postgres_ai on the path to Self-Driving Postgres. First: tackling the problem most teams actually face: index maintenance, not index creation. Sad truth: everyone agrees when we show what needs to be done. But it's too much effort, so it sits in TODO forever. Until DB hits a wall 1/

there are performance cliffs in Postgres, there are just cliffs -- they are not less painful to fall from excited about new postgres.ai product we're building now – it will help predict / detect / mitigate all possible cliffs

Deep dive in the situation with prepared statements and partitioned tables in Postgres: part 1 postgres.ai/blog/2025102... part 2 postgres.ai/blog/2025102... part 3 postgres.ai/blog/2025103...

this repo deserves more stars gitlab.com/postgres-ai/... we aim to have most detailed and advanced monitoring for Postgres -- and it's FOSS because it's just a foundation of something bigger (after 20 years with Postgres, next month will be the most exciting announcement of my career)

postgres_ai monitoring 0.11: VictoriaMetrics, lots of optimizations, new dashboards for table and index analysis, and more github.com/postgres-ai/...

postgres_ai monitoring 0.11: VictoriaMetrics, lots of optimizations, new dashboards for table and index analysis, and more github.com/postgres-ai/...

TIL during this hacking session: REFRESH MATERIALIZED VIEW CONCURRENTLY creates temp tables github.com/postgres/pos... – so, if you do it super frequently, it might contribute to system catalog bloat

live session of Postgres [vibe] hacking happening now -- join! www.youtube.com/watch?v=6cZv...

I'm on a trip until 2025-10-22, and have only limited time with laptop – #PostgresMarathon will be continued when I'm back

#PostgresMarathon 2-008: LWLock:LockManager and prepared statements postgres.ai/blog/2025101...

#PostgresMarathon 2-007: Should we worry about pg_blocking_pids()'s observer effect? postgres.ai/blog/2025101...

Got very much attracted by title "Boosting Planning Performance | Scaling Postgres 387", was very curious (because I do want to learn more ways to boost planning performance), checked it out: www.youtube.com/watch?v=d7__... ...and ERROR: infinite recursion detected

Postgres hacking session LIVE www.youtube.com/watch?v=5lN7...

#PostgresMarathon 2-001 To warm up, let's talk about lightweight and heavyweight locks (or "regular locks" or just "locks"). I'm using these materials: - PG docs (first of all, www.postgresql.org/docs/current...) 1/

#PostgresMarathon is back! Make sure you follow me so we could learn some good stuff about Postgres together. This time, I'll be doing it a bit differently: - I'll choose one big Postgres topic that I'll be covering over a week or two 1/

anti-tip: SELECT INTO is ...DDL. It creates a table. Not what you expect after spending hours inside pl/pgsql www.postgresql.org/docs/current... I wish Postgres didn't have this feature 😂

If you think that you understand the whole meaning of max_locks_per_transaction, think again

What do you like most in Postgres 18? Look at www.postgresql.org/docs/release... and tell me. I'll go first, my top-3 list:

Some PG18 posts spread misinformation.. Be careful, we see a lot of LLM-generated BS lately.

this needs to be in PG core github.com/fabriziomell... or this github.com/munakoiso/lo...

If you deal with index bloat, and ready to try out a new thing – reach out to me, we're have something really cool

just merged: WAL-G now got Prometheus exporter github.com/wal-g/wal-g/...

levels of psql-fu: level 0: psql <connoptions> -c "SELECT 1;" level 1: pg_isready <connoptions> level 1: psql -c "select;" # conn. options in env vars ... level 99: psql -XAatcselect 2>&1 | ts | tee -a log

Canceling Disney+ subscription is a good idea today. It also has a "why" text field that is worth filling.

How do you define "bloat" in Postgres? Do you include dead tuples into bloat? Or only empty space (reclaimed and tracked by FSM, after VACUUM deletes dead tuples)?

[vibe]hacking Postgres online with Andrey, Kirk www.youtube.com/watch?v=mVYl... – it's live, join!