Repository navigation
Fix oversized OTLP metric exports in direct and buffered delivery.
- Split snapshots into requests of at most 256 points, with an additional conservative estimate for histogram expansion.
- Preserve complete histograms, values, cumulative start times and observation times across requests.
- Keep metric spool entries separate. On a retryable failure, retain only unaccepted and unsent requests in their original order.
Upgrade every spool drainer as well as the application. Already queued oversized snapshots are not repartitioned. A single indivisible oversized point can still be rejected by the receiver, and existing spool capacity/crash-loss semantics remain unchanged. The batching budget does not raise the receiver's total-series limit.
Validation: PHP/Laravel CI matrix and Redis integration passed; 1,374 local tests including benchmarks passed with the repository's 512 MiB test budget. Audit, dependency-license and SBOM checks passed. All 7,831 points in an actual ID snapshot reconstructed exactly after batching.