Rates Engine v0.5.0-rc.53

[v0.5.0-rc.53] — 2026-05-15

Fixed

• Coverage snapshot now populates DURING an all-time backfill.
  /v1/diagnostics/ingestion's backfill_coverage was built from
  the background trades-scan cache (BackfillCoverageStats). Under
  the all-time SDEX backfill that query is too IO-contended to
  finish within its timeout, so the cache stayed empty and the
  status page showed "Coverage snapshot pending" indefinitely —
  exactly when operators most need it. Rebuilt cursor-first:
  density_pct / covered / expected / earliest / latest for every
  on-chain source now derive purely from the union of completed
  backfill-cursor intervals (no trades scan), so the snapshot is
  live even mid-backfill. The trades-scan cache is demoted to
  best-effort trade_count enrichment + off-chain CEX/FX context
  rows; an empty or stale cache can no longer blank the whole
  section. earliest_ledger / latest_ledger are now the
  merged-cursor processed span (honest "what we walked"), not a
  trades MIN/MAX that could imply an interior gap is covered —
  density_pct remains the gap-aware number. Added
  soroswap-router + defindex to sourceGenesisLedger so the
  two new on-chain Soroban sources get an honest density bar
  instead of falling through to the un-mapped cache-only path.

Changed

• max_locks_per_transaction 256 → 4096 (archival-node ansible
  role default + R1 postgresql.conf). The per-transaction lock-table
  sizing knob drives the instance-wide lock table
  (max_locks_per_transaction × max_connections). TimescaleDB takes
  one lock per chunk a query scans — not just INSERTs; any broad
  SELECT over the now-2,738-chunk trades hypertable
  (per-source coverage, the pools/markets CTE, cagg/fx coverage)
  locks one entry per scanned chunk. The 64→256 hand-bump from the
  2026-05-06 SEV-3 lasted ~9 days before the table grew enough that
  concurrent diagnostics + the all-time SDEX backfill thrashed it
  again (a single query observed holding 26,927 locks; coverage
  snapshot stuck "pending", /v1/pools?order_by=pair 500s, cagg/fx
  coverage out of shared memory). Re-sized to 4096 (819,200-entry
  table) for permanent growth headroom rather than another
  incremental bump — memory cost ≈220 MB, negligible against the
  48 GB shared_buffers / 188 GB host. Applied live to R1 (postgres
  restart) + codified in the role so R1 rebuilds and R2/R3 cutover
  inherit it.