Re: Compression of bigger WAL records

Hi Andrey,

I've captured some perf of a backend doing
CHECKPOINT;
SELECT count(*) FROM (SELECT * FROM pgbench_accounts FOR UPDATE);
in a loop to check the difference with and without reusing zstd cctx
for 0001. There's definitely a visible improvement with less time
spent in ZSTD_compress_using_Dict.

+ * Compression context reused across all block images compressed by this
+ * backend. zstd keeps its match tables and window in here, roughly 1.3MB at
+ * the default level, and allocates them on first use. Creating a context per
+ * call would repeat that allocation for every full-page image.

I'm a bit surprised by the 1.3MB estimation, how did you find this
amount? If I check with ZSTD_sizeof_CCtx just after the context is
created, I have:
(gdb) p ZSTD_sizeof_CCtx(zstd_cctx)
$1 = 5288

After being used once, it seems to stay stable with:
(gdb) p ZSTD_sizeof_CCtx(zstd_cctx)
$5 = 154261
So it looks closer to 152KB.

+ if (zstd_cctx == NULL)
+ zstd_cctx = ZSTD_createCCtx();

Would it be worth it to register ZSTD_freeCCtx using on_proc_exit? I
imagine this will be tagged as leaked memory since it's never freed.

Regards,
Anthonin Bonnefoy