From: melancholictheory (@melancholictheory) <noreply+melancholictheory@github.com>
To: pgbackrest/pgbackrest <noreply+pgbackrest-pgbackrest@github.com>
Subject: Re: [pgbackrest/pgbackrest] PR #2912: Limit the free list to the configured I/O buffer size
Date: Mon, 09 Nov 2026 12:11:48 +0000
Message-ID: <gh-pgbackrest-pgbackrest-2912-c6739740033@github.com> (raw)
In-Reply-To: <gh-pgbackrest-pgbackrest-2912@github.com>
References: <gh-pgbackrest-pgbackrest-2912@github.com>
I tried the workload described in #2816 on Linux ARM64 with local MinIO. So far that's the only platform I've tested for this workload.
I created 10,000 small PostgreSQL 18 tables, took one full backup with block incremental enabled, and updated every other table. Each build then ran six differential backups against the same full backup, with the build order rotated between rounds. The runs used process-max=1, buffer-size=1MiB, and zstd level 1.
| Build | Median elapsed time | Median minor faults | Median peak summed RssAnon |
| --- | ---: | ---: | ---: |
| #2914 revert | 3.22 s | 302k | 28.3 MiB |
| Original free list | 2.96 s | 24k | 50.8 MiB |
| Single bucket for buffer-size | 2.88 s | 22k | 34.6 MiB |
The single bucket was faster than the revert in five of six rounds. The median difference was about 11%, or 6% when excluding the first round, where the revert took 8.38 s. That gives us a modest elapsed-time improvement on this ARM64 workload alongside a much larger reduction in minor faults. The x86-64/KPTI case is still untested.