SYSTEM Cited by 1 source
StatsD¶
StatsD is the canonical line-oriented metrics protocol (originally from Etsy, 2011) that clients emit over UDP/TCP to a local aggregator, which then flushes aggregated counters/timers/gauges to a downstream metrics backend. Its strength is simplicity and universal client libraries; its classic deployment is as a sidecar next to the application.
On Lyft / LyftLearn¶
Lyft's Kubernetes-era ML workloads emitted StatsD metrics to a per-pod sidecar container — the classic deployment shape (see patterns/sidecar-agent). When Lyft's ML platform moved the compute half onto SageMaker, SageMaker has no sidecar support — the sidecar assumption broke.
The fix, baked into Lyft's cross-platform base images, was to reconfigure the in-process StatsD client and the networking layer to connect directly to Lyft's metrics aggregation gateway rather than a local sidecar. User-facing metrics emission API was unchanged (Source: sources/2025-11-18-lyft-lyftlearn-evolution-rethinking-ml-platform-architecture).
This is a concrete instance where the managed platform's lack of sidecars forces the aggregation tier to move from per-pod to cluster/region-level, with the transport pushed up the stack.