Benchmarks

What an SDK costs. Measured, not promised.

Two questions a generated SDK has to answer: how much of your bundle does importing one operation drag in, and how much CPU does calling it burn once the network is out of the picture. These are the numbers, straight from the harnesses in benches/.

Bundle size

Import one operation, pay for one operation.

Each row is a small worker that imports a single Distilled operation and calls it, bundled the way Alchemy bundles for Cloudflare Workers. Deep and barrel imports produce the same bytes; the two rows marked ^* import more than that on purpose.

118.4 KBgzipped, one S3 operation

61.7 KBgzipped, one Workers operation

129.7 KBgzipped, S3 + Workers together

aws-s3-deep@distilled.cloud/aws/s3 + /Credentials, one op (GetObject)
118.4 KB gzip410.7 KB raw222 modules
1/112 aws/s3no leaks
aws-s3-deep+nopure@distilled.cloud/aws/s3 + /Credentials, one op (GetObject)
118.9 KB gzip412.3 KB raw222 modules
1/112 aws/s3no leaks
aws-barrel@distilled.cloud/aws root barrel (+ /s3 deep) — same op
118.7 KB gzip410.7 KB raw222 modules
1/112 aws/s3no leaks
aws-services-index^*@distilled.cloud/aws/index — all ~430 services as namespaces, only S3.getObject used
118.4 KB gzip410.7 KB raw222 modules
1/112 aws/s3no leaks
cf-workers-deep@distilled.cloud/cloudflare/workers + /Credentials, one paginated op (listScripts)
61.7 KB gzip222.2 KB raw106 modules
1/74 cloudflare/workersno leaks
cf-workers-deep+nopure@distilled.cloud/cloudflare/workers + /Credentials, one paginated op (listScripts)
62.6 KB gzip225.2 KB raw106 modules
1/74 cloudflare/workersno leaks
cf-barrel@distilled.cloud/cloudflare root barrel (Services.* — ~120 services), same op
61.7 KB gzip222.2 KB raw106 modules
1/74 cloudflare/workersno leaks
combined-worker^*alchemy-worker-shaped: s3.getObject + workers.listScripts (two deep imports)
129.7 KB gzip464.8 KB raw234 modules
1/112 aws/s31/74 cloudflare/workersno leaks

^* aws-services-index imports all ~430 AWS services as namespaces; combined-worker calls two operations, one per provider. Every other row imports and calls exactly one.

rolldown 1.2.5 · bun 1.3.13 · Intel(R) Core(TM) Ultra X7 358H · 3 runs · 2026-09-09 · adc438102

Runtime

Per call, with the network removed.

The HTTP client is mocked, so this is only the SDK's own work: encode, sign, serialize, parse, decode. Lower is better, so rows run fastest p50 first; the faint bar on Full call rows is the mocked round-trip with no SDK at all.

10 µsmock round-trip, no SDK

31 µscloudflare createNamespace · p50

442 µsaws GetCallerIdentity · p50 (SigV4)

operation	p50	p99
cloudflare`kv.createNamespace`3-field envelope	31 µs	43 µs
cloudflare`r2.createBucket`5-field envelope	34 µs	36 µs
cloudflare`d1.createDatabase`7-field envelope	37 µs	39 µs
cloudflare`workers.putScriptSecret`EnvelopePayloadRoot response	37 µs	242 µs
cloudflare`kv.listNamespaces`paginated protocol, 20-namespace envelope	40 µs	268 µs
cloudflare`d1.listDatabases`paginated protocol, 20-db envelope	49 µs	247 µs
cloudflare`zones.createZone`~30-field envelope	51 µs	314 µs
cloudflare`r2.listBuckets`20-bucket envelope	51 µs	292 µs
cloudflare`zones.listZones`paginated protocol, 20-zone envelope	273 µs	856 µs
aws`s3.HeadObject`rest-xml + SigV4, header-only 200	374 µs	4.7 ms
cloudflare`workers.listScripts`paginated protocol, 20-script envelope	401 µs	1.2 ms
aws`sts.GetCallerIdentity`aws-query + SigV4, XML result	442 µs	2.4 ms
aws`lambda.Invoke`rest-json + SigV4 (signed payload), streamed response	547 µs	3.5 ms
aws`dynamodb.GetItem`aws-json + SigV4, 7-attr item	547 µs	3.5 ms
aws`dynamodb.PutItem`aws-json + SigV4, 7-attr item	597 µs	3.7 ms
aws`lambda.GetFunction`rest-json + SigV4, full config body	685 µs	3.5 ms
aws`s3.ListBuckets`rest-xml + SigV4, 20-bucket XML body	912 µs	5.0 ms

bun 1.3.13 · Intel(R) Core(TM) Ultra X7 358H · quick profile · 2026-09-09 · 9b3f5da1d

Method — how this is measured

Bundle. A tiny worker-shaped entry imports one Distilled operation and calls it, bundled with rolldown using the same options Alchemy uses for Cloudflare Workers: bun resolve conditions, minified single ESM chunk, PURE annotations on. Size is the chunk; gzip is what the wire sees. Tree-shake quality is the number of operations that survive minification out of the service's total, with a check that no other service or provider leaked in.
Runtime. The HttpClient is replaced by one that returns a canned response, so what remains is the SDK's own work: schema encode, request build (including SigV4 for AWS), response parse, envelope decode, retry wrapper. Numbers are per call, measured with mitata on Bun. The baseline row is the mocked round-trip with no SDK in the loop; subtract it to see the SDK's share.
Caveats. One developer machine, quick profile. Bytes are stable; timings move ±30% between runs and machines. Nothing here is compared against a vendor SDK yet. Treat it as a shape, not a guarantee.
Reproduce: pnpm bench:bundle and pnpm bench:runtime from the repo root.