Alex Georgiev

PostgreSQL 19 RC1 can replicate sequence values to a logical subscriber, closing a long-standing failover gap. I measured it: an instant match on subscribe, then a frozen value through 700 more replicated rows until I ran REFRESH SEQUENCES, and a promotion I did properly still hit the exact duplicate-key error the feature is meant to prevent.

FastAPI 0.162.0 stopped silently turning every omitted boolean form field into False, including ones declared Form(...) as required. I measured the before and after on four field shapes, and found an unrelated sibling-parameter break that shipped the same week.

Redis 8.12 RC2 adds per-member TTLs to sets. A membership check dropped from roughly 585 to 290 microseconds over 2,000 calls, but setting the TTL itself wasn't faster, and converting a small intset cost 15 times more memory.

SQLite 3.54.1 cut a 50,000-row no-op UPDATE from 2,143 WAL pages to 511, exactly matching a table with no expression index at all. Once the indexed value actually changes, the saving disappears entirely.

I benchmarked TypeScript 7's Go-based compiler against 6.0.3 on a real NestJS checkout: 2.5x to 4.5x faster, not Microsoft's claimed 8-12x, and 7.0's stable release used more memory than the old compiler while dropping the API ts-node needs.

Bun 1.4.4 ships a built-in formatter. Checking Prettier's own 2,623-file source tree took 0.10-0.15s under bun fmt against 36.77s under Prettier itself, but it also silently reimplements only a short whitelist of Prettier plugins.

I measured Deno's new deno task caching: 3.9s cold against 9ms warm on a real build step, but a 3000-file input tree made the cache check itself about 4x slower than running an untracked task.

I measured Django 6.1's new MAILERS config system against the EMAIL_BACKEND it deprecates: 50 emails cost 50 SMTP connections either way, and only a manual trick (unrelated to MAILERS) got it down to 1.

PostgreSQL 19 adds ON CONFLICT DO SELECT. I measured a get-or-create call drop from two round trips to one under contention, found it gives no benefit when there's no conflict, and checked what it refuses, costs and locks.

I measured Python 3.14's new subinterpreters against threads and processes: 3.6s of GIL contention dropped to 0.9s, but a single interpreter flatly refuses a second concurrent thread.

Valkey 9.2.0-rc1's whole-stream LZ4 RDB compression shrank a million-row JSON dataset from about 178MB to 35MB, an 80% cut the old per-string LZF default barely touches. It does nothing for replication unless a second, undocumented-in-the-release-notes setting is also flipped.

A Caddy reload that broke my running server used to come back HTTP 200. On 2.11.6 and 2.11.7 it's a 400 with a real error, and a related fix cuts a stuck Unix socket from about two minutes to instant.

I opened nginx's new control API over plain TCP with no credentials and used it to read live worker PIDs and swap a running config from outside the container. The unix-socket default stopped a non-owning user cold, reload itself gained no speed, and just having the listener open cost about 300KB of resident memory.

I benchmarked Git's new last-modified command against a per-file git log loop on git/git: 861ms versus 2981ms listing one real directory, but slower once the same 40 paths were scattered across the tree, and a parent-then-child pathspec order silently drops a row.

On a cgroup v1 host, Kubernetes 1.35 and 1.37's kubelet crash-looped 82 times in 90 seconds instead of booting; the documented fix stops the crash but needs an undocumented second flag for kubeadm, and quietly breaks Memory QoS.

I reproduced a Docker Engine 29.8.2 bug where docker save followed by docker rmi left 100% of a 4.5MB image's layers on disk, confirmed 29.9.1 fixes it, and found no prune command could see the leak on the broken version.

Redis 8.4 adds IFEQ, IFNE, IFDEQ and IFDNE to SET. I measured a 1MB compare-and-set drop from about 4.3ms to 1.6ms using the new digest option, then found the official Python client silently drops the server's own GET reply on one code path.

SQLite 3.54 replaces sqlite3_analyzer with a built-in diskused() function. It ran a disk-space report in 0.025s against the old tool's 0.061s, but on a build without SQLITE_ENABLE_DBSTAT_VTAB it fails with no error and no output at all.

Docker Compose 5.6 added a jobs: section for one-off tasks. up, create and start all refuse to touch one, a scheduled trigger makes up fail outright, and only docker compose run actually starts a job, paying about 15% more when its dependency has to start cold.

I measured PostgreSQL 19 beta4's new parallel autovacuum on a six-index, 400,000-row table: 4 workers launched every time, but wall-clock vacuum time stayed within 4% of the serial run, while dead-tuple memory doubled.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle ...

I benchmarked DuckDB 2.0 alpha's new triggers on a 500,000-row insert: row-level triggers cost about 3x a plain insert, statement-level ones with transition tables cost about 1.3x. Half the documented trigger matrix isn't implemented yet.

I bloated a Postgres 19 beta table, then rewrote it with VACUUM FULL and with the new REPACK CONCURRENTLY: 4.3-5.9s of blocked client time against 0.3-0.9s, and a documented replication-slot limit I managed to actually hit.

I tested deno audit on three old npm packages: it found 13 real GitHub advisories in under 0.3s, but --fix, --ignore and --frozen-lockfile all failed to do what their own docs describe.

I tested Traefik 3.7's new service-level middlewares on a 90/10 weighted canary service: 20 of 200 requests landed on the canary branch and all 20 carried the canary-only header, with zero leaks across 500 concurrent requests. Traefik's own announcement blog shows YAML that doesn't parse on the real software.

I benchmarked MongoDB 8.3's new arrayIndexAs against the range-and-arrayElemAt workaround it replaces: 134ms versus 102ms server-side at 320,000 elements, a gap that shrinks to about 5% once ten clients compete for the same CPU cores.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass ...

Node.js 26 enables the Temporal API with no flag. I found it fixes a real one-hour DST drift in naive Date arithmetic, but costs about 19 times more than Date per operation, and its default overflow mode doesn't throw where I expected it to.

Django 6.1 raised the default PBKDF2 iteration count from 1,200,000 to 1,500,000. I measured the extra cost per login, then found it rewrites every hash on next login, including ones an admin had deliberately hardened above the new default.

I benchmarked pg_search 0.25.11 against 0.26.0 on the same three-million-row table: a ten-term OR search dropped from 129ms to 29ms, but a single-term search got five times slower.