Under your OTLP exporter config, add a sending_queue block; set enabled: true (default), queue_size to the maximum number of batches to buffer in memory (default 1000), and num_consumers to the number of goroutines dequeuing batches (default 10)
To survive Collector restarts without data loss, add a file_storage extension and set storage: file_storage in the sending_queue block; this persists the queue to disk between restarts
Add a retry_on_failure block; set enabled: true, initial_interval to the first retry backoff (e.g., 5s), max_interval to the ceiling of the exponential backoff (e.g., 30s), and max_elapsed_time to the total time to keep retrying before giving up (e.g., 10m)
Set queue_size based on expected data rate and acceptable downtime: queue_size × average_batch_size_bytes must fit within available memory; for a 10 000 spans/s pipeline with 512-span batches, 1000 queue entries buffer roughly 50 seconds of data
Monitor otelcol_exporter_queue_size and otelcol_exporter_queue_capacity to detect sustained backlog; set an alert when queue_size / queue_capacity exceeds 0.8 so you can scale the backend before the queue fills and drops begin
Combine with the batch processor (placed before the exporter) to reduce the number of queue entries and RPCs; larger batches mean fewer queue slots consumed per unit of telemetry volume
Known gotchas
The in-memory queue does not survive Collector crashes without a file_storage extension; data buffered at crash time is lost—enable file_storage when export durability is required
max_elapsed_time counts from the first failed send attempt; if the backend is down for longer than max_elapsed_time, buffered data is dropped even if the queue has capacity—balance max_elapsed_time against acceptable data loss window
num_consumers controls concurrency on the queue; increasing it improves throughput to backends that support parallel connections (gRPC) but can overwhelm backends with per-connection rate limits—benchmark before increasing above the default
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp