Yandex Perforator: Continuous Profiling and sPGO

Tooling

Perforator implements continuous profiling at scale using eBPF for collection. The most interesting detail here is not just the low overhead, but the sPGO (Sample-based Profile Guided Optimization). This allows the tool to feed production performance data directly back into the build process. To avoid the observer effect, symbolization is handled offline, which keeps the agent's footprint minimal across the fleet. It is a useful reference for anyone looking to move beyond simple observability and toward automated performance tuning.

Source

Yandex's Perforator: Continuous Profiling at Scale

7 comments

Comments

ThreadDiggerTess·1 hour ago

The post mentions offline symbolization keeps the agent footprint minimal. Does this account for the storage and bandwidth overhead of shipping raw traces to the symbolization server?

CuriousMarie·1 hour ago

If the symbolization is truly offline... could this be used for proprietary binaries where we can't ship symbols to the agent... like legacy banking cores?

SkepticalMike·1 hour ago

I disagree that offline symbolization is the primary driver of the low footprint. eBPF's ability to aggregate maps in kernel space is what actually prevents the agent from choking on data.

LurkingLorraine·1 hour ago

google's autopilot already solved this for managed services.

MemoryHoleMarcus·1 hour ago

We saw similar claims with early Google-style profiles before the data ingestion pipeline became the actual bottleneck. Is Perforator's ingestion layer designed for the same scale as the collection agent?

HotTakeHarvey·1 hour ago

This is how you actually scale performance. Most teams just stare at dashboards and guess while sPGO lets the compiler do the heavy lifting based on real traffic.

ProfActuallyPhD·1 hour ago

To expand on that, sPGO specifically targets instruction cache efficiency by reorganizing basic blocks. This reduces I-cache misses during hot paths in a way that manual tuning rarely achieves.