Repository navigation
Fixes
- Postgres cursor reads are fast again. Since 1.0.1 each page's query looked so expensive to the planner that Postgres JIT-compiled it every time, making a page several times slower than its rows. Which rows are read is unchanged. On older releases,
ALTER ROLE <reader> SET jit = offavoids the cost. - An ordered sink gets nothing past a failed batch. With
concurrencyabove 1, batches read behind a failed send to afileorobject_storeoutput were still written, so the retried batch landed after them. They are now handed back unsent. - Python and Node: a drain ends with a custom middleware on the input. A middleware registered with
register_middleware/registerMiddlewaredid not passexit_on_emptyto the source, so the route never exited. - Python and Node: a custom middleware keeps the endpoint's ordering, flush and hooks. Behind a middleware on the output, a
filesink withconcurrencyabove 1 could write batches out of order, and commits could be released before a flush-durable sink flushed.
Performance
- Python handlers: handler threads keep their Python thread state, and
Caller.call()/ JSON handlers encode faster. An HTTP route answering from a Python handler went from ~78k to ~87k req/s at 8 connections. parallelandbufferno longer clone every message's outcome when settling a batch (about a third of the CPU on a batched lookup behind aCaller).filter:number(field)compared with a number parses only that field instead of running the expression engine on a JSON value.
What's Changed
See CHANGELOG.md for details.
Full Changelog: v1.6.0...v1.6.1