Rates Engine v0.5.0-rc.62

[v0.5.0-rc.62] — 2026-05-20

Fixed

• Prewarm extended to all 7 readers fired by /v1/assets/{id}
  (full deferred-#37). rc.61 partial covered only
  GetCoinByAssetID per verified asset; live measurement on r1
  post-rc.61 confirmed the canonical-form path still dropped to
  2.2s on first hit because the other SIX readers fired by the
  handler (GetCoinTopMarkets(id, 5),
  GetCoinPriceHistory24h, GetCoinPriceHistory7d,
  GetCoinMarketsCount, GetCoinTradeCount24h, GetCoinATH)
  cold-filled per request. Subsequent hits served sub-ms only
  because the first request populated all 7 SWR slots. Full fix:
  new prewarmAssetDetail(ctx, logger, coins, assetID) helper
  that calls EVERY one of the 7 readers per verified canonical
  asset_id PLUS native. Limit=5 on GetCoinTopMarkets matches
  the handler's literal — per the
  feedback_prewarm_handler_drift memory, any drift in
  args (limit, order, sources) means a different SWR cache key
  → silent miss → the same bug class. Each reader's prewarm
  call logs at Debug; transient failures don't block subsequent
  readers. Net effect post-deploy: every verified-currency
  canonical-form lookup (and native) should land sub-200ms on
  FIRST hit, not just subsequent.
  (#34 residual).** Postgres can issue SQLSTATE 57014
  (canceling statement due to user request) from server-side
  statement_timeout / lock_timeout /
  idle_in_transaction_session_timeout — none of which trip
  clientAborted (the http request context is alive) or
  handlerTimedOut (the per-call context hasn't deadlined). The
  error reached the issuers handler as a bare 57014 and fell
  through to a 500. Combined with the sla-probe's threshold
  set at exactly 99.0%, a single transient per ~430-sample
  burst put availability under threshold and fired
  ratesengine_sla_probe_unit_failed_alert. Fix: new
  transientStorageErr(err) helper in envelope.go that
  classifies the three transient classes (SQLSTATE 57014 not
  carried by context cancellation, driver: bad connection
  after pool retries exhausted, and EOF / broken-pipe network
  blips) and returns 503 with the issuers-transient problem
  type. Standard handler ordering preserved: clientAborted →
  handlerTimedOut → transientStorageErr → 500. The 503 still
  registers as a non-2xx for the sla-probe's availability
  metric, BUT — operator runbook: configure the probe to
  count 5xx-but-not-503 as failures, OR loosen the threshold
  to 98.5%. The cleaner long-term path is operator-facing.
• Density formula no longer over-credits via interior-gap
  bridging (user-reported). Status-page reported 100% density
  for soroswap-router and defindex while the #38 historical
  backfill was only ~78% through their range. Root cause:
  extendWithLiveTail (diagnostics_ingestion.go:1056) was
  bridging interior gaps between two backfill intervals whenever
  the upper bracket's start ≤ liveTop, on the assumption that
  live ingest had walked the gap. That assumption is FALSE for
  sources added to enabled_sources after live ingest had
  already crossed the gap-end ledger — exactly the case for
  soroswap-router + defindex which were enabled at rc.5x while
  the live cursor was already at ~62.5M; the interior
  [60M, 62.5M] gap got false credit. Fix: remove interior-gap
  bridging entirely. Live-tail credit is now head-band only
  — from the top of the backfill union up to min(liveTop, tip).
  The previously-protected edge case ("disjoint high
  gap-backfill island silently capping density at ~96%") becomes
  honest under-coverage: operators close such gaps with a
  targeted backfill rather than silent live-credit. Density on
  defindex/router should now drop from the false 100% to the
  honest ~78%-and-rising as #38 progresses. Two existing tests
  that locked in the bridging behaviour were updated to the new
  honesty-first policy.