Commit c93080b0a3bf4faffb375169c5ec06adce05eb2f | the-knife

author	Theo Browne <me@t3.gg> 2026-10-07 01:32:35-07:00
committer	GitHub <noreply@github.com> 2026-10-07 01:32:35-07:00
log	c93080b0a3bf4faffb375169c5ec06adce05eb2f
tree	1941163151a11f7ba1ec19fe4c2776835d325117
parent	10395bcc59707302c1476e19186532ff5a9d3557
signature	Signed by PGP key B5690EEEBB952194

perf(server): background branch lookups share one GitHub query per sweep (#16760)

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

2 files changed, 69 insertions(+), 15 deletions(-)

apps
- server
  - src
    sourceControl
    GitHubCli.test.ts
    GitHubCli.ts

apps/server/src/sourceControl/GitHubCli.test.ts+53-12

...	@@ -251,7 +251,7 @@ describe("GitHubCli repository resolution", () => {	...	@@ -251,7 +251,7 @@ describe("GitHubCli repository resolution", () => {
251	describe("GitHubCli.listPullRequestsByHead", () => {	251	describe("GitHubCli.listPullRequestsByHead", () => {
252	const remotes = remotesOutput(["origin", "git@github.com:acme/web.git"]);	252	const remotes = remotesOutput(["origin", "git@github.com:acme/web.git"]);
253		253
254	it.effect("reads heads on one repository in one GraphQL document", () => {	254	it.effect("reads a background sweep's staggered heads in one GraphQL document", () => {
255	const documents: Array<GitHubApi.GitHubGraphQlInput> = [];	255	const documents: Array<GitHubApi.GitHubGraphQlInput> = [];
256	const { layer } = harness({	256	const { layer } = harness({
257	remotes,	257	remotes,
...	@@ -267,20 +267,23 @@ describe("GitHubCli.listPullRequestsByHead", () => {	...	@@ -267,20 +267,23 @@ describe("GitHubCli.listPullRequestsByHead", () => {
267	});	267	});
268	return Effect.gen(function* () {	268	return Effect.gen(function* () {
269	const gh = yield* GitHubCli.GitHubCli;	269	const gh = yield* GitHubCli.GitHubCli;
270	const lookups = yield* Effect.all(	270	const lookup = (headSelector: string) =>
271	["feature/a", "feature/b"].map((headSelector) =>	271	gh
272	gh.listPullRequestsByHead({	272	.listPullRequestsByHead({
273	cwd: "/repo",	273	cwd: "/repo",
274	headSelector,	274	headSelector,
275	state: "all",	275	state: "all",
276	limit: 100,	276	limit: 100,
277	rateLimitHost: "github.com",	277	rateLimitHost: "github.com",
278	}),	278	})
279	),	279	.pipe(Effect.forkChild);
280	{ concurrency: "unbounded" },	280	// Each branch's own git reads come first, so a sweep's lookups arrive spread out.
281	).pipe(Effect.forkChild);	281	const firstLookup = yield* lookup("feature/a");
282	yield* TestClock.adjust("50 millis");	282	yield* TestClock.adjust("200 millis");
283	const [first, second] = yield* Fiber.join(lookups);	283	const secondLookup = yield* lookup("feature/b");
		284	yield* TestClock.adjust("300 millis");
		285	const first = yield* Fiber.join(firstLookup);
		286	const second = yield* Fiber.join(secondLookup);
284	assert.deepStrictEqual(	287	assert.deepStrictEqual(
285	first?.map((pr) => pr.number),	288	first?.map((pr) => pr.number),
286	[7],	289	[7],
...	@@ -298,6 +301,44 @@ describe("GitHubCli.listPullRequestsByHead", () => {	...	@@ -298,6 +301,44 @@ describe("GitHubCli.listPullRequestsByHead", () => {
298	}).pipe(Effect.provide(layer));	301	}).pipe(Effect.provide(layer));
299	});	302	});
300		303
		304	it.effect("caps a background document at twenty-five heads", () => {
		305	const headCounts: Array<number> = [];
		306	const { layer } = harness({
		307	remotes,
		308	api: {
		309	graphql: (input) =>
		310	Effect.sync(() => {
		311	const heads = Object.keys(input.variables ?? {}).filter((key) => /^h\d+$/.test(key));
		312	headCounts.push(heads.length);
		313	return encodeJson({
		314	data: { repository: Object.fromEntries(heads.map((key) => [key, { nodes: [] }])) },
		315	});
		316	}),
		317	},
		318	});
		319	return Effect.gen(function* () {
		320	const gh = yield* GitHubCli.GitHubCli;
		321	const lookups = yield* Effect.all(
		322	Array.from({ length: 26 }, (_, index) =>
		323	gh.listPullRequestsByHead({
		324	cwd: "/repo",
		325	headSelector: `feature/${index}`,
		326	state: "all",
		327	limit: 100,
		328	rateLimitHost: "github.com",
		329	}),
		330	),
		331	{ concurrency: "unbounded" },
		332	).pipe(Effect.forkChild);
		333	yield* TestClock.adjust("500 millis");
		334	yield* Fiber.join(lookups);
		335	assert.deepStrictEqual(
		336	headCounts.toSorted((a, b) => a - b),
		337	[1, 25],
		338	);
		339	}).pipe(Effect.provide(layer));
		340	});
		341
301	it.effect("matches an owner:branch selector on the head owner", () => {	342	it.effect("matches an owner:branch selector on the head owner", () => {
302	const { layer } = harness({	343	const { layer } = harness({
303	remotes,	344	remotes,
...	@@ -355,10 +396,10 @@ describe("GitHubCli.listPullRequestsByHead", () => {	...	@@ -355,10 +396,10 @@ describe("GitHubCli.listPullRequestsByHead", () => {
355	.listPullRequestsByHead({ cwd: "/repo", headSelector, state: "open", limit: 1 })	396	.listPullRequestsByHead({ cwd: "/repo", headSelector, state: "open", limit: 1 })
356	.pipe(Effect.flip, Effect.forkChild);	397	.pipe(Effect.flip, Effect.forkChild);
357	const missing = yield* read("missing");	398	const missing = yield* read("missing");
358	yield* TestClock.adjust("50 millis");	399	yield* TestClock.adjust("500 millis");
359	assert.strictEqual((yield* Fiber.join(missing))._tag, "GitHubCliUnavailableError");	400	assert.strictEqual((yield* Fiber.join(missing))._tag, "GitHubCliUnavailableError");
360	const limited = yield* read("limited");	401	const limited = yield* read("limited");
361	yield* TestClock.adjust("50 millis");	402	yield* TestClock.adjust("500 millis");
362	const error = yield* Fiber.join(limited);	403	const error = yield* Fiber.join(limited);
363	assert.strictEqual(error._tag, "GitHubCliRateLimitError");	404	assert.strictEqual(error._tag, "GitHubCliRateLimitError");
364	assert.propertyVal(error, "retryAt", 123);	405	assert.propertyVal(error, "retryAt", 123);

apps/server/src/sourceControl/GitHubCli.ts+16-3

 @@ -352,8 +352,17 @@ const HEAD_LOOKUPS_PER_DOCUMENT = 50;
 /**
  * How long a head lookup waits for company. Branch discovery reaches GitHub only after each
  * branch's own git reads, so lookups started together arrive tens of milliseconds apart.
+ * A background sweep's lookups spread over up to ~300ms, and every document costs a point no
+ * matter how few heads it holds, so reads without reserve wait longer.
  */
 const HEAD_LOOKUP_BATCH_WINDOW = "50 millis";
+const BACKGROUND_HEAD_LOOKUP_BATCH_WINDOW = "500 millis";
+/**
+ * Background documents fill up under the longer window, and a failed document fails every head
+ * in it. Fifty `main`-like heads of a hundred pull requests each took up to ~10s, GitHub's own
+ * processing limit; twenty-five took ~7s.
+ */
+const BACKGROUND_HEAD_LOOKUPS_PER_DOCUMENT = 25;
 /** A full document is 5,000 rows of well under 2 KB each. */
 const HEAD_LOOKUP_MAX_RESPONSE_BYTES = 16_000_000;
 /**
 @@ -705,9 +714,13 @@ export const make = Effect.gen(function* () {
         ),
       );
     },
-  }).pipe(
+  }).pipe(RequestResolver.batchN(HEAD_LOOKUPS_PER_DOCUMENT));
+  const interactiveHeadResolver = headResolver.pipe(
     RequestResolver.setDelay(HEAD_LOOKUP_BATCH_WINDOW),
-    RequestResolver.batchN(HEAD_LOOKUPS_PER_DOCUMENT),
+  );
+  const backgroundHeadResolver = headResolver.pipe(
+    RequestResolver.batchN(BACKGROUND_HEAD_LOOKUPS_PER_DOCUMENT),
+    RequestResolver.setDelay(BACKGROUND_HEAD_LOOKUP_BATCH_WINDOW),
   );
   const listByHead = Effect.fn("GitHubCli.listByHead")(function* (input: {
 @@ -733,7 +746,7 @@ export const make = Effect.gen(function* () {
         limit: ownerMatch ? OWNER_HEAD_SCAN_LIMIT : limit,
         allowReserve: input.allowReserve,
       }),
-      headResolver,
+      input.allowReserve ? interactiveHeadResolver : backgroundHeadResolver,
     );
     if (!ownerMatch) return rows;
     const headOwner = ownerMatch[1]!.toLowerCase();