Azure/azure-sdk-for-go sdk/data/azcosmos/v1.5.0-beta.7

on GitHub

1.5.0-beta.7 (2026-06-02)

Features Added

• Added retry policy for transient 500, 502, and 504 server errors on read requests. The request is retried once in the current region and, if applicable, once against the next preferred region. Writes are not retried. This matches the behavior of the .NET, Java, and Python Cosmos SDKs. See PR 26821.

Bugs Fixed

• Fixed missing OTel tracing spans for internal queries executed by ReadManyItems. Each per-partition query page now creates a query_items span, matching the tracing behavior of NewQueryItemsPager. See PR 26813.
• 403/WriteForbidden retries refresh the global endpoint manager fire-and-forget (CAS-gated) instead of blocking on a synchronous gem.Update. See PR 26889.
• Connection-error retry policy now attempts up to 3 retries against the current region before failing over, and performs at most one cross-region failover per call. Cross-region failover for writes only occurs when the error proves the request never reached the service (DNS, dial, TLS handshake, ECONNREFUSED, etc.); writes on ambiguous transport failures (e.g. ECONNRESET, EOF, transport-level timeouts) no longer fail over to another region, avoiding potential duplicate writes. Reads still fail over for any transport error. Caller-set context deadlines or cancellations short-circuit the policy without consuming the caller's budget with retries. See PR 26858 and PR 26915.
• HTTP 408 Request Timeout responses are now handled by the Cosmos client retry policy: reads are retried exactly once against another region, and writes are returned to the caller immediately to avoid potential duplicates. See PR 26858.
• Fixed excessive GetDatabaseAccount HTTP calls when using preferred regions, and stopped data-plane retries from trailing into the customer-supplied (default) endpoint once account topology is populated. See PR 26815.
• Partition key range cache now serves concurrent callers from a single in-flight refresh per container, and the cached routing map remains readable while a refresh is in progress. The refresh runs on a detached background context.Background() so a caller's cancellation no longer aborts the shared fetch for other waiters; each caller continues to honor its own context deadline. See PR 26855.
• Partition key range cache change-feed pagination is now resilient to mid-drain throttling. 429 responses are retried indefinitely (with capped linear backoff + jitter) since the service is explicitly asking the client to slow down, and the pages already accumulated are preserved instead of restarting the drain from page 1 on the next refresh. See PR 26855.

Other Changes

• Tightened the default HTTP client: 5s dial timeout (down from azcore's 30s), 65s http.Client.Timeout wall-clock cap per HTTP attempt (was unbounded), larger idle connection pool (1000 total / 100 per host, up from azcore's 100 / 10), and faster HTTP/2 health checks. Caller-supplied Transport and shorter context deadlines are unaffected. See PR 26856.