Infrastructure
Measured on Postgres 16 at one million rows: jsonb costs 2.3x on disk and 1.4x on point lookups, both survivable. The number that actually hurts is the update path — one index on a jsonb column kills every HOT update and multiplies WAL by 7.8x, even when the update never touches the indexed field.
AI & Data
Yes. int8 scalar quantisation makes a vector set 3.96x smaller and returned 0.984 of the float32 top-10 at 384 dimensions — 1.000 once the top 100 candidates were re-scored with float32. What it does not do is make the scan faster. Measured storage, scan speed and rank agreement at 384 and 1,536 dim
Integration
408, 425, 429, 500, 502, 503 and 504 are worth retrying; 400, 401, 403, 404 and 422 never are. The hard part is the network errors and the POSTs — a timeout on a POST means the request may already have succeeded.
Infrastructure
Alpine saves 86 MB over slim and every classic reason to avoid it has quietly expired. What still costs you is sharp running 1.8x slower on musl, and apk. Measured on Node 22.
Integration
We benchmarked a SKIP LOCKED job queue on Postgres 16 at 1 to 100 concurrent workers. It peaks near 16,000 claims a second — and the first thing to break is autovacuum, not row locks.
AI & Data
Brute-force cosine similarity over vectors in SQLite stays under 100 ms to roughly 300,000 rows at 384 dimensions — if the dot product is not running in a Python loop, where the same wall arrives seventeen times earlier. Measured latency, storage and what an HNSW index actually buys.
Infrastructure
Measured on Postgres 16 at one million rows: a UUID primary key costs about 2.6x on bulk insert, 1.4x on index size, and nothing at all on cached point reads. Time-ordered UUIDv7 removes most of the write cost — and made the index bigger, not smaller, which turned out to be the interesting part.
Infrastructure
A /health that always returns 200 tells you the HTTP listener works — which was never the thing that broke. Split liveness from readiness, put a timeout on every dependency check, and never let a downstream failure trigger a restart that cannot fix it.
Infrastructure
Shipping software customers run themselves means every install is a machine you cannot log into. Five things — a bounded support window, idempotent migrations, a preflight that names the missing variable, a curlable /health, and a support bundle — decide whether that is a product line or a queue you
AI & Data
Split on structure first — headings, paragraphs, tables — and fall back to character count only when a single block is genuinely too big. Fixed-size chunking cuts tables in half and hands your retriever a fragment that no embedding model can repair.
AI & Data
SQLite FTS5 gives you ranked, snippet-annotated, as-you-type search with no server and no network. The parts that catch people out are the external-content pattern, the delete trigger, and what the default tokenizer does to accented text.
Integration
Conditional requests turned 60 polls into one unit of quota, and a keyset cursor stopped the poller silently dropping records that shared a timestamp. The three bugs are all invisible — nothing errors, you just quietly lose rows.