mirror of
https://github.com/DeBrosOfficial/orama.git
synced 2026-08-06 18:29:28 +00:00
dc7abbcc51
Three anchat-priority fixes for delivery-critical paths. #137 — rarely-invoked function WASM cold-fetch timed out (15s) → INVOKE_FAILED with no retryable flag: - uploadWASM now pins WASM on EVERY cluster peer (replication=-1) so no gateway node is ever cold, and HARD-FAILS the deploy on a pin error (no silent "deployed but unfetchable"). - GetWASMBytes fetches with an independent 4s×3 budget decoupled from the function's timeout, returning the typed ErrWASMFetchTimeout. - New FUNCTION_UNAVAILABLE rpc code (retryable) emitted on BOTH the HTTP and WS invoke error frames; the invoker no longer re-runs (re-fetches) on it. #139 — a gateway's local rqlite follower can lag minutes; none-reads silently served stale rows and recipients missed messages: - New in-process freshness gate (LocalFollowerFresh via raft /status last_contact + commit/applied gap, TTL-cached). A stale local follower's none-reads AUTO-DEGRADE to the leader-routed weak conn — transparent, no caller change, never stale. #138 — expose rqlite native level=none&freshness=<dur> in db_query_batch: - New queryNoneFresh custom HTTP path (gorqlite can't express freshness) + typed ErrFreshnessViolation; freshness field on the db_query_batch envelope, validated/clamped at the boundary; a violation returns a structured stale_rejected envelope so WASM can fall back to a leader read. Reads-only; additive + backward compatible (empty freshness = prior behavior). Verified live on devnet: rqlite v8.43.0 signals a violation as HTTP 200 + {"error":"stale read"}; /status is reachable unauthenticated for the gate.