[receiver] high resource utilization during the time of uploading TSDB blocks about thanos HOT 8 CLOSED

It looks like snappy compression is using way more cpu for some reason in 0.33

Edit: Its because snappy compression for gRPC was added in 0.28, you could try 0.33 and use

--receive.grpc-compression=none

and see if it helps? Another hint is that your bandwidth with 0.26 seems to be orders of magnitude higher, also pointing at snappy compression.

See:

      --receive.grpc-compression=snappy
                                 Compression algorithm to use for gRPC requests
                                 to other receivers. Must be one of: snappy,
                                 none

from thanos.

Confirming that the use of the --receive.grpc-compression=none field in version 0.33.0 was effective, and now the resource utilization is similar to that in version 0.26.0. Thanks @MichaHoffmann!

from thanos.

@hayk96 I don't think you need it there. The routers will take care of firing all the requests to achieve replication. The ingesters have nothing more to do than read the incoming non-compressed request and write the data.

from thanos.

I have mixed opinions regarding this. CPU is usually cheaper than network bandwidth. Ingest latency is super important too. If to save network bandwidth we need to pay in latency, this is a very sensible decision. Latency with the compression in @hayk96's environment was awful compared to no-compression. And check that memory usage chart...

from thanos.

I uploaded the profiles to pprof.me:

heap profiles:

• v0.33: https://pprof.me/3a91f164ab7ce91989b90bc866ea0ad2
• v0.26: https://pprof.me/be0808d689c3ce00fb2c655e7ab30955

cpu profiles

• v0.33: https://pprof.me/3105bcf0b4bd4d9cbfda27640ba23851
• v0.26: https://pprof.me/74cdccfc18754305f15af03c37c6fc50

from thanos.

BTW should I add that flag in the receive-ingesters as well, does it make sense?

from thanos.

Maybe we add this in the Troubleshooting; Common cases? What do you think?

from thanos.

Maybe we add this in the Troubleshooting; Common cases? What do you think?

I'm indifferent i think; Its a bit of a tradeoff between network bandwidth and CPU; and usually CPU is cheaper so i think defaulting to snappy here is the correct behavior. But then again you faced the regression during update and would have profited from an FAQ article about it. I dont know whats the best course of action tbh!

from thanos.