Cloudflare adds on-demand production profiling for Workers and Durable Objects

In the dashboard the profiler sits under Build > Compute > Workers & Pages. It is on each Worker's Observability tab, behind a "Flamegraph" option in a dropdown ^[14]. You choose a duration and a version, press Capture Profile, and the graph renders once the profiler has run for that long ^[4][6]. Each rectangle is a function call. Its width is the CPU time or memory that function used ^[6]. The CLI does the same in one command. The documented example passes `--duration-ms 5000` and `--profile-type cpu` for the latest version and redirects the output to a file ^[3].

The announcement covers Durable Objects alongside Workers ^[1]. Every step in the walkthrough, including the CLI's `--worker-id` flag, is written for a Worker ^[3][14].

The R2 example is the most useful part of the post. Cloudflare profiled the Worker behind the R2 binding, a high-traffic service, for 50 seconds ^[8]. The widest boxes were decryptBlock, which decrypts object data on the way out, and fillResponse, which moves bytes ^[9]. Cloudflare found no clear way to optimize either ^[9]. A storage service should be spending its CPU on those two jobs. The lead came from the table view sorted by samples, which Cloudflare recommends because some boxes are hard to see in the graph ^[10]. The second candidate was a duplicate call to metrics ^[13].

The 2.7x is a function-level figure. A function holding 5% of CPU that runs 2.7x faster drops to about 1.9% of the original total. That saves about 3.1% of the Worker's sampled CPU time ^[15]. The post puts the function at over 5%, so 3.1% is a floor ^[15]. Cloudflare wrote that on this Worker "eliminating even the smallest amount of wasted CPU cycles can have a massive impact on its performance." ^[8]

Two conditions decide whether that gain transfers to another team's Worker. The code needs the same mistake: a replacer that walks the tree itself while JSON.stringify is already calling it on every node ^[11]. The payloads also need depth. In Cloudflare's case a value nested five levels deep was processed five times, so passes per value tracked nesting depth ^[11]. A Worker serializing flat objects would recover little from the same fix ^[16].

The post does not say what a capture costs in overhead on a live Worker, how often the profiler samples, or which plans include it.