I must say I'm quite pleased to see how well Go version works. It does only use 1.5x the CPU and (predictably) much more RAM/VRAM, but not a crazy amount either (the expected increase is 2x).
Of course you can write a more optimal version in C / C++ / Zig / Rust, but at the same time Go is much easier to write and you don't pay for the convenience with an absurd performance loss like in Python or PHP
tbf it took 4 years since PG15 support was added for me to fix remote BASE_BACKUP support & wal-g base backups being inconsistent on PG15+ (parameter typo had pg_backup_stop return before wal archived far enough for consistency)
The importance of optimizing (resource) margins and having predictable memory usage increases significantly in the DBaaS/Postgres world, where your process coexists and competes with other critical workloads.
Also, WAL-RUS isn’t rocket science. Postgres already exposes a bunch native constructs for WAL archival, making development fairly straightforward even in Rust.
TL;DR: when to choose Rust or Go really depends on the workload and what you are going after.
That coexistence is also why the GC-free rewrite helps more than the speed numbers suggest. An archiver is allocation-light until it hits a compression burst, then the Go heap can spike toward 2x live right when Postgres wants that memory. GOMEMLIMIT caps the spike but pays in GC CPU during exactly those bursts, so on a small instance, you are trading OOM risk for throughput. Rust removes that dial.
As someone who only has a cursory knowledge of Postgres backup systems, how does this compare to something like pgBackRest? When would someone reach for one over the other?
If you're running pg yourself I recommend pgbackrest. It doesn't run as a daemon, & it forks multiple processes for concurrency. But it's simple to run as archive_command & is light on resources outside concurrency
Do you have any benchmarks with a mix of long open transactions and short ones? I've struggled a lot with WAL-E in the past there, and am curious if that changes here.
no. but wal-g & wal-rus both have parallelism over wal-e. however are you more asking about handling build up of wal / vacuum prevention caused by long running transactions? those are up to postgres, archive command only keeps pushing wal so that when postgres is ready to get rid of wal it can. seems like your scenario wouldn't care much what the archiver is since wal should be shipped long before postgres is ready to get rid of wal
Quick one on the benchmark: was the 2.8GB peak virtual or resident? Go reserves a large virtual arena it mostly never faults in, so RSS tends to be a fraction of the virtual peak, and if Postgres headroom was getting squeezed off the virtual number you were sizing against memory the kernel never actually charges for.
Correct. We tune overcommit so postgres reliably returns out of memory. It becomes complicated to accurately tune overcommit for every AWS instance type. We configure GOMEMLIMIT/cgroups but those are about RSS. Outliers come together: instances running queries out of memory on our service tend to also be pushing other resource limits, causing wal-g & prometheus exporters to start having more erratic memory usage at the worst time
This helps on both ends of the cost spectrum. Large 64 core instances are where our heuristics fall off the most as variance increases, & tiny instances with 8GB of memory can use every 100MB of RSS we can get
You probably could limit the bloating of Go programs by setting GOMAXPROCS to something like 1 or 2 on smaller machines, but then again you wouldn't get the best performance. So IMO good call here to rewrite it in a language without GC.
The article’s distinction about memory is a bit too simplified. Virtual memory is address space mapped or reserved by the process. RSS is resident physical pages, not necessarily “actively used.”. RSS also overcounts shared memory, which is why PSS/USS matter.
I must say I'm quite pleased to see how well Go version works. It does only use 1.5x the CPU and (predictably) much more RAM/VRAM, but not a crazy amount either (the expected increase is 2x).
Of course you can write a more optimal version in C / C++ / Zig / Rust, but at the same time Go is much easier to write and you don't pay for the convenience with an absurd performance loss like in Python or PHP
indeed, wal-g actually started as a port of wal-e which was Python: https://www.citusdata.com/blog/2017/08/18/introducing-wal-g-...
wal-g was a much larger improvement over wal-e. we're optimizing the margins here
I'd like to use the one getting the most community support. So too soon to wait for Rust vs Go. Although on paper, Rust is better.
tbf it took 4 years since PG15 support was added for me to fix remote BASE_BACKUP support & wal-g base backups being inconsistent on PG15+ (parameter typo had pg_backup_stop return before wal archived far enough for consistency)
https://github.com/wal-g/wal-g/pull/2262
but yes, this is young project, so fair take
++ We’re big Go fans, most of PeerDB is written in Go: https://github.com/PeerDB-io/peerdb
The importance of optimizing (resource) margins and having predictable memory usage increases significantly in the DBaaS/Postgres world, where your process coexists and competes with other critical workloads.
Also, WAL-RUS isn’t rocket science. Postgres already exposes a bunch native constructs for WAL archival, making development fairly straightforward even in Rust.
TL;DR: when to choose Rust or Go really depends on the workload and what you are going after.
That coexistence is also why the GC-free rewrite helps more than the speed numbers suggest. An archiver is allocation-light until it hits a compression burst, then the Go heap can spike toward 2x live right when Postgres wants that memory. GOMEMLIMIT caps the spike but pays in GC CPU during exactly those bursts, so on a small instance, you are trading OOM risk for throughput. Rust removes that dial.
PostgreSQL WAL-RUS, no relation to PostgreSQL WALRUS https://github.com/supabase/walrus
As someone who only has a cursory knowledge of Postgres backup systems, how does this compare to something like pgBackRest? When would someone reach for one over the other?
If you're running pg yourself I recommend pgbackrest. It doesn't run as a daemon, & it forks multiple processes for concurrency. But it's simple to run as archive_command & is light on resources outside concurrency
wal-g/wal-rus have higher throughput
Do you have any benchmarks with a mix of long open transactions and short ones? I've struggled a lot with WAL-E in the past there, and am curious if that changes here.
no. but wal-g & wal-rus both have parallelism over wal-e. however are you more asking about handling build up of wal / vacuum prevention caused by long running transactions? those are up to postgres, archive command only keeps pushing wal so that when postgres is ready to get rid of wal it can. seems like your scenario wouldn't care much what the archiver is since wal should be shipped long before postgres is ready to get rid of wal
Yeah, I'm probably misremembering some details there. Thanks
How to Design an index layer for postgres WAL-G backups to make a paging VFS like sqlite-http-vfs for pglite in WASM?
Great name
Quick one on the benchmark: was the 2.8GB peak virtual or resident? Go reserves a large virtual arena it mostly never faults in, so RSS tends to be a fraction of the virtual peak, and if Postgres headroom was getting squeezed off the virtual number you were sizing against memory the kernel never actually charges for.
Correct. We tune overcommit so postgres reliably returns out of memory. It becomes complicated to accurately tune overcommit for every AWS instance type. We configure GOMEMLIMIT/cgroups but those are about RSS. Outliers come together: instances running queries out of memory on our service tend to also be pushing other resource limits, causing wal-g & prometheus exporters to start having more erratic memory usage at the worst time
This helps on both ends of the cost spectrum. Large 64 core instances are where our heuristics fall off the most as variance increases, & tiny instances with 8GB of memory can use every 100MB of RSS we can get
You probably could limit the bloating of Go programs by setting GOMAXPROCS to something like 1 or 2 on smaller machines, but then again you wouldn't get the best performance. So IMO good call here to rewrite it in a language without GC.
The article’s distinction about memory is a bit too simplified. Virtual memory is address space mapped or reserved by the process. RSS is resident physical pages, not necessarily “actively used.”. RSS also overcounts shared memory, which is why PSS/USS matter.