TxHash ingestion + query benchmarks (both backends)

[karthikiyer56]

TL;DR

Benchmark both ingestion and query paths for the txhash store across both backends — RecSplit-backed (cold: #696 writer + cold reader) and RocksDB-backed (hot store, read + write) — to confirm we hit ingestion-throughput and query-latency SLAs across the full history. Symmetric to Tamir's ledger benchmarks (#726).

What to measure

Ingestion throughput

• RecSplit (cold): txhashes / sec via the four-phase pipeline (COUNT, ADD, BUILD, VERIFY) at production-shaped tx-index dataset sizes.
• RocksDB (hot): txhashes / sec via PutTxHash calls from a simulated live-ingestion harness.

Query latency

• Cold backend (RecSplit) only: P50 / P90 / P95 / P99 for Lookup(txhash) across the 16 CFs.
• Hot backend (RocksDB) only: P50 / P90 / P95 / P99 for Lookup(txhash) across the 16 CFs.
• Federated reader (hot + cold): P50 / P90 / P95 / P99 with a representative production hot/cold mix.

Resource footprint

• Memory and disk footprint at representative dataset sizes (e.g., 1 yr, 5 yr, full chain).

Acceptance

• Benchmark report comparing both backends across all metrics above, runnable from a documented harness in the repo.
• Federated-reader latency reflects the hot/cold mix expected in production.
• Reproducible: the harness seeds the data, runs the benchmarks, emits structured numbers (CSV or JSON) for plotting.
• Report identifies any bottleneck above the agreed SLA, with a triage note (e.g., "needs index tuning", "needs CF tuning", "out of scope for this milestone").

Dependencies

• Blocked by the cold reader sub-issue.
• Blocked by the hot store sub-issue.
• Tracked under the txhash store epic (cross-cutting per-entity view).
• Native parent: streaming epic Streaming epic — orchestration + live ingestion in the unified RPC v2 service #704 (real-work home).

[tamirms]

Subsumed by #771 (full-history benchmark commands — the single home for ingest + query benchmarks across all data types). Txhash ingest/query benchmarking is covered there.