Łukasz Mierzwa

I can move it here - https://github.com/prometheus/prometheus/blob/707600d84f5567354ef48d4593ddab84c4837811/tsdb/db.go#L967 But one thing to consider here, I think, is that mmapping of chunks was added to lower memory usage, so if we mmap...

I wonder if mmap should also be forced in `ChunkSnapshot()` - https://github.com/prometheus/prometheus/blob/9558b9b54bd3d0cb1d63b9084f8cbcda6b0d72fb/tsdb/head_wal.go#L581

Deployed latest patch to another instance and it also show improvement. No longer seeing rule evaluation duration spikes leading to missed iterations. Memory usage looks similar to vanilla v2.35.0 ![Screenshot...

> I see that you have already done some investigation in #10377. Did you have the queue enabled? Did this patch with the queue make the situation better? Enabling chunk_write_queue...

> I see that you have already done some investigation in #10377. Did you have the queue enabled? Did this patch with the queue make the situation better? I *think*...

One of the most difficult aspects of operating Prometheus is staying on top of resource usage, which mostly means memory. The problem is that if Prometheus scrapes too many metrics...

What alternatives would you consider @roidelapluie ?

``` tsdb/head.go:1512:16: fieldalignment: struct with 224 pointer bytes could be 72 (govet) type memSeries struct { ``` For my biggest instances with 25M time series having 25M memSeries would take...

Weirdly `unsafe.Sizeof()` reports same size before and after reordering fields. It should include padding in the result (https://go-review.googlesource.com/c/go/+/397516) so either there's no loss to padding and govet checks are incorrect...

> Maybe there is a difference in 32 bit systems Possibly, there's https://github.com/golang/go/issues/47055 that I found. Although I'm getting same (seemingly inaccurate) results with https://pkg.go.dev/golang.org/x/tools/go/analysis/passes/fieldalignment as the ones reported with...