Benchmark: Gitaly with Reftables and Transactions

Objective

Benchmark Gitaly with Reftables+ Transactions enabled.

Scope

• Baseline: Gitaly
• Variant: Gitaly with Reftables + Transactions enabled
• Workloads: clone/fetch, push/write (small/large), RPC hot paths, GC/pack, repo create/delete, mixed read/write, concurrency scaling (repos × RPCs)
• EXT4 FS

Configuration:

Inside config.yml this needs to be declared for each gitaly instance that needs txn enabled

 config:
      # Whether transactions should be enabled.
      transactions: true

To enable reftables for a specific repo, initialize the test repo within the repositories block in config.yml with

reference_backend: reftable

Test Matrix

• Concurrency:
  • Repos: 1...5
  • Parallel RPCs per repo: 1, 8, 32
• Repository shapes:
  • Small repos (many loose objects/files)
  • Medium repos (Decent number of commits, trees, branches)
  • Large monorepos (large packfiles, deep history)
• Transactions Enabled
• Reftables Enabled/Disabled
• Cache state:
  • Cold-cache (drop_caches) vs steady-state warm cache
    • by default we use cold-cache to calculate worst-case performance

Documentation Requirements

• Exact Gitaly, Git, kernel versions

Deliverables

• Reproducible benchmark runs using the framework from #6732
• Comparison tables/plots: baseline vs Transactions (per workload and concurrency level)
• Analysis summarizing:
  • bottlenecks when lots of refs are used
  • p95 latency per RPC
• Artifact bundle per run: raw results (JSON/CSV), system profile, flamegraphs
• Summary published to the Gitaly Wiki with links to artifacts