Pinterest has joined forces with Intel to expand gProfiler, an open-source continuous profiler that delivers production performance visibility across internet-scale infrastructure. The social platform runs hundreds of thousands of instances serving 640 million monthly active users, with services built in Java, Python, C++, Go, and Node.js. At that scale, even minor efficiency gains can translate into substantial infrastructure and cost savings, according to a report published by InfoQ.
The challenge Pinterest faced was fragmented tooling. Engineers relied on separate instruments like perf, async-profiler, and py-spy, generating disconnected profiles that proved difficult to correlate across programming languages and machines. Instead of constructing a proprietary profiling platform, Pinterest collaborated with Intel to enhance gProfiler and return those improvements to the open-source community. The partnership introduced dynamic profiling, hardware-level telemetry, wall-time profiling, and performance-counter flame graphs, while reinforcing the system for deployment across large production fleets. Agents typically remain dormant, consuming under 1% CPU and 100 MB of memory, and receive profiling instructions via periodic heartbeats. The new capability lets engineers select services, hosts, or clusters and activate profiling remotely through the UI or API, transforming profiling into an operational function rather than something requiring shell access to individual machines.
The report notes that gProfiler functions as an orchestration layer over multiple existing profilers, merging system-wide perf data with runtime-specific instruments such as async-profiler for Java and eBPF-based profiling for Python to produce unified flame graphs across processes running on a host. This matters because modern services often span multiple layers—a Java application might spend time inside the JVM, native libraries, or the Linux kernel, while a Python service can involve interpreter, native, and kernel activity. The platform separates collection from analysis: agents operate on individual hosts, while gProfiler Performance Studio aggregates and visualizes resulting profiles, storing them in S3, passing them through SQS for asynchronous indexing, and warehousing them in ClickHouse for analytical queries. According to the authors, Pinterest and Intel also hardened the platform for production, addressing concerns around agent resource consumption, disk usage, and multi-terabyte ingestion before deploying it across the fleet.
The distinction with gProfiler lies in its emphasis on system-wide, multi-runtime profiling and open-source extensibility, the report explains. Rather than replacing instruments like perf or async-profiler, it orchestrates them and adds hardware-level profiling through Intel's tooling. Integration with Intel PerfSpect allows gProfiler to gather hardware-level information such as cache behavior, memory bandwidth, and CPU stalls, while Pinterest added PMC flame graphs that correlate functions with hardware performance events like cache misses and branch mispredictions. Wall-time profiling makes waiting periods visible—a service can be slow while consuming relatively little CPU because threads are waiting on I/O, locks, or remote procedure calls, which CPU profiling alone won't reveal. At small scale, profiling is typically something an engineer enables when a performance problem surfaces, but at Pinterest's scale, the cost of waiting for engineers to reproduce issues or manually profile machines becomes prohibitive.
The report concludes that continuous profiling provides a baseline, while dynamic profiling allows teams to investigate specific anomalies when they occur. By combining runtime profiles, system behavior, and hardware performance data under an open-source platform, Pinterest is effectively making profiling another layer of production observability. gProfiler sits in an increasingly crowded continuous-profiling landscape alongside Parca, Grafana Pyroscope, and commercial platforms like Datadog, but its focus on multi-runtime orchestration and hardware telemetry sets it apart. Pinterest specifically notes that Intel's VTune provides much deeper processor and application performance analysis, but it was designed primarily as a performance-engineering tool rather than a cloud-native fleet profiler, which influenced the decision to build on gProfiler rather than adapt VTune for its environment. The open-source contribution means other organizations running heterogeneous, large-scale fleets can now access the same visibility without building proprietary tooling. For enterprises balancing developer productivity against infrastructure spending, this shift positions profiling as a continuous feedback loop rather than a reactive debugging step, though success will hinge on whether engineering teams invest in the cultural change required to act on the data routinely.