A fix for #9447 Before this change, for the workload in the ticket, we saw a memory stay at around 4 GB an hour after the workload ended. after this change we see both a lower memory high watermark, and also see memory come back to baseline at the end as we expect (after the `HASURA_GRAPHQL_PG_CONN_LIFETIME` and a subsequent `_idleStaleReaperThread` interval elapses). But note the numbers given on any given machine may be slightly different, since the behavior depends on the version of glibc and processor speed (Since the test case depends on the server being overloaded). This might also help some users commenting in https://github.com/hasura/graphql-engine/issues/9592 PR-URL: https://github.com/hasura/graphql-engine-mono/pull/9823 GitOrigin-RevId: 5650aa42d10d46c418c21686983a982d69011884 |
||
---|---|---|
.. | ||
mimalloc | ||
README.md | ||
update-sources.sh |
Vendored mimalloc
This is mimalloc vendored for use in
our server code. You can upgrade mimalloc by editing VERSION
in
update-sources.sh
and re-running it.
Usage sites in haskell projects look like:
-- See 'preload-mimalloc/README.md'
-- We want this in all server binary executable sections
common preload-mimalloc
c-sources: ../preload-mimalloc/mimalloc/src/static.c
include-dirs: ../preload-mimalloc/mimalloc/include
cc-options: -DMI_MALLOC_OVERRIDE
Usage in haskell projects
In #9447 we see overloading the server with large and slow queries results in high memory usage and what looks like a space leak, involving foreign data from libpq.
See that ticket for details, but mimalloc helps (relative to my glibc) by:
- better avoiding fragmentation, resulting in lower overall memory usage (4GB peak 2.8 aftervs.
- eagerly decomitting memory so that OS-reported memory usage is a useful
metric (we do the same in the haskell RTS by specifying
--disable-delayed-os-memory-return
)
We also benefit by standardizing this, rather than relying on users' glibc version, which might vary in behavior, making debugging difficult.
Other implementations such as tcmalloc or jemalloc were not tested but might perform better or be more tunable.
Why not LD_PRELOAD?
It's most common and simplest to use the dynamic linker to override malloc. This would work in docker containers we ship, but we want all users to have the same (better) behavior out of the box.