Vector Panda mascot, a panda in an orange hoodie and beanie, holding a card-catalog box of indexed posts

Build log

Architecture decisions, performance dives, and the gotchas behind shipping a distributed vector database.

July 2026

7 posts

June 2026

5 posts

May 2026

4 posts

April 2026

4 posts

March 2026

18 posts

metadata filter predicates

MongoDB-style filters on query API, wired through the whole stack.

mutation design reversal

Threw out sidecar mutation files, went with in-place shard mutation.

warm tier is live

Warm tier using mmap over NFS, faster than expected.

hosted E2E tests

Real E2E tests on .150, one collection per format, exact top-k validation.

collection status field

Added explicit status field instead of inferring readiness from internal state.

delete was broken everywhere

Delete broken in 4 places at once, had to implement two-phase deletion.

JSONL duplicate keys

Duplicate keys under parallel JSONL load, two bugs that looked like one.

key collisions

Generated keys colliding in multi-file collections.

bvecs seek bug

Binary vector readers weren't rewinding between reads, garbage dimensions.

safetensors ID bug

Was ignoring the explicit ID tensor in safetensors files.

csv dimension columns

Convention for CSV dimension columns, plus Feather zstd.

format explosion

Added six vector formats in an afternoon, every one broke ID handling.

meta.idx + meta.dat

Binary metadata index, under 5ms for 500 lookups.

metadata is slow

Returning metadata in query results, first two approaches too slow.

gateway topology

Consumer site is the only public entry point, everything else is internal.

python SDK

First commit of the Python SDK, points at the gateway.

self-service signup

Self-service signup and API keys, no more manual account creation.

killed the base fee

Free tier on warm storage, renamed cold to paused, 512D normalization for metering.

January 2026

4 posts

August 2025

5 posts

July 2025

5 posts

June 2025

16 posts