Epic: optimize WAL ingest pipeline

[erikgrinaker]

Motivation

We should investigate WAL ingestion and optimize for end-to-end throughput. In particular for migrations of new users, typically pg_restore or logical replication.

DoD

Ingestion should saturate hardware (either disk or network IO) on at least one of compute, safekeeper, or pageserver.

Implementation ideas

TBD

Tasks

Tasks

Beta Give feedback

1. Investigate WAL ingest performance for logical messages #9642
   a/performance c/storage +2
   
2. Investigate WAL ingest performance for bulk writes #9789
   a/performance c/storage +2
   
3. Investigatepg_restore performance #9623
   a/benchmark a/performance c/storage +3
   
4. remote_storage: revisit throttling/ratelimiting #3698
   c/storage/pageserver +1
   
5. remote_storage: add request/response size histograms #9819
   a/observability c/storage +2
   
6. pageserver: slow get_relsize() during WAL ingestion with many relations #9855
   a/performance c/storage/pageserver +2
7. safekeeper: optimize writing of zeroed segments #9662
   a/performance c/storage/safekeeper +2
   
8. safekeeper: avoid or optimize segment fsyncs #9664
   a/performance c/storage/safekeeper +2
   
9. safekeeper: optimize control file fsyncs on commit_lsn advancement #9663
   a/performance c/storage/safekeeper +2
   
10. safekeeper: eagerly return AppendResponse with flush_lsn on segment fsync #9688
    a/performance c/storage/safekeeper +2
    
11. CPU/heap profiling at runtime #9534
    a/observability c/storage +2
    
12. safekeeper: increase segment size #9687
    a/performance c/storage/safekeeper +2
    
13. safekeeper: use tokio-epoll-uring for file IO #9741
    a/performance c/storage/safekeeper +2
14. safekeeper: improve batching of pipelined AppendRequest #9689
    a/performance c/storage/safekeeper +2
    
15. safekeeper: return AppendResponse on txn commit #9690
    a/performance c/storage/safekeeper +2
    

Other related tasks and epics

• Investigate bulk load performance #1263