Time travel is just a really good index
Why AS OF queries on 30 days of history cost less than one extra replica.
Customers use AS OF queries to answer the question every incident review asks: what did the system believe at 14:02? Keeping 30 days of history sounds expensive. It costs us less than a single extra replica.
Versions, not snapshots
Basalt never overwrites a row. Each change appends a new version tagged with the log sequence that produced it. A query AS OF a timestamp maps that time to a sequence and reads the newest version at or below it.
The index does the work
Versions of one key sit next to each other on disk, newest first.
A sparse index stores one entry per 4 KB block, so seeking to a version is one block read.
Old versions compact into columnar files after 24 hours and compress about 9 to 1.
The result: a point read 30 days back takes 6.1 ms at p99, against 4.2 ms for the present. History adds 14% to storage on a typical workspace.
More from the engine room.
How we cut p99 from 11 ms to 4.2 ms
Three changes to the read path in v4.0: declared sort keys, a per-core allocator and a planner that stopped re-sorting partitions at query time.
Exactly-once without two-phase commit
Idempotent sinks, epoch fencing and a 64-bit sequence per partition. No coordinator, no stalls.
Orbitly deleted 14 pipelines in a quarter
How a fintech team replaced nightly batch jobs with six materialized views.
