Ulrik Falklöf

Thanks @mostynb! We have patched bazel client with a custom load balancer. It distributes load over several bazel-remote instances. The patch contains logic to ignore cache instances for a minute...

Sorry for later answer, took time to get hardware and to experiment. When trying #639, bazel-remote rejects requests instead of crashing, which is good. But it rejected requests even when...

Interesting! What do you have in mind for implementing “_asynchronously at a slower rate_”? Can it avoid creating a goroutine for each blob? Example: If using 5% of a 2...

> then we could allow immediate removal for small blobs I'm striving to perform evictions with **low priority** (compared to the handling of incoming requests) rather than immediate removal of...

Thanks @AlessandroPatti! Status update from me: I did several experiments and surprisingly found that deleting lots of files **sequentially with a single goroutine**, is much faster than starting separate go...

Most benchmarks I have been running are re-sending traffic patterns from recorded real sessions. I need to think about if I can share something from those. But for cache eviction...

As reference with same pre-filled cache of 10 GB tiny files: ``` $ time rm -rf cache real 0m41.752s user 0m0.693s sys 0m40.106s ``` A naive attempt to use parallellism...

Thanks for running benchmark @AlessandroPatti! Both #695 and #696 are introducing queues for evicted files, and I think both our benchmarks motivates that approach. I think next step is to...

See also https://github.com/buchgr/bazel-remote/pull/696/commits/81961893732633db52e2ed2214229bc21f120dfa from https://github.com/buchgr/bazel-remote/pull/696 that introduces bazel_remote_disk_cache_size_bytes_limit with the labels "evict" and "reject".

I believe (but I'm not 100% sure, and it could be related to my internal patches) that "PUT ... unexpected EOF" is logged by bazel-remote, when bazel aborts an upload...