Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

test(proxy): pin the 768 MiB heap cap and the 2048Mi limit for the compact worst-structure load

Alfredo Lopez996d19e87ca3
diff --git a/codex-llm-proxy/test/deployManifest.test.ts b/codex-llm-proxy/test/deployManifest.test.ts
index 968cffafc..06650a6f3 100644
--- a/codex-llm-proxy/test/deployManifest.test.ts
+++ b/codex-llm-proxy/test/deployManifest.test.ts
@@ -19,31 +19,31 @@ const CONTROL_PLANE_CONFIG = readFileSync(
/** Headroom between the last in-flight request finishing and SIGKILL. */
const SHUTDOWN_MARGIN_SECONDS = 20
/**
- * Heap cap the peak below was measured with. At 384 MiB the process aborts
- * while it parses the three admitted ordinary bodies when each one carries the
- * worst structure the contract admits (262144 containers, 262144 members,
- * 1048576 elements), before any stream reaches the upstream. 512 MiB is the
- * smallest cap that survives that load.
+ * Heap cap the peak below was measured with. The byte-based BodyBudget admits
+ * about eight compact bodies of the worst structure the contract admits
+ * (262144 containers, 262144 members, 1048576 elements; 2.8 MiB each), and each
+ * parsed tree costs about 44 MiB after a collection. With 8 held streams, 8
+ * queued bodies and two visual streams, a 640 MiB cap aborted the proxy in 3 of
+ * 3 runs; 768 MiB passed in 3 of 3. 768 MiB is the smallest cap tested that
+ * passes, not a proven minimum.
*/
-const HEAP_CAP_MIB = 512
+const HEAP_CAP_MIB = 768
/**
- * The largest measured peak, heap capped at HEAP_CAP_MIB (tsc build, one
- * process, upstream request through undici, macOS, Node v24.18.0):
- * - D5 (#739): eight 8 MiB streams, two 24 MiB visual streams and three
- * queued 8 MiB bodies with a text tool description: 790.3 MiB at a
- * 384 MiB cap;
- * - the same load with every ordinary body carrying the worst structure the
- * contract admits (#806 Q1): 1094.9 and 1129.2 MiB in two runs at 512.
- * That is the recorded peak.
- * Measured by hand, not by this suite (harness: 8 held streams, 3 queued
- * bodies, 2 visual bodies). Re-measure before changing it.
+ * The largest measured peak RSS, heap capped at HEAP_CAP_MIB (real proxy built
+ * with tsc, upstream request through undici, macOS, Node v24.18.0, the compact
+ * worst-structure shape: 8 held streams, 8 queued bodies, two visual streams).
+ * Three runs: 1445.5, 1459.5 and 1327.3 MiB. The recorded peak is the maximum,
+ * because the run-to-run spread on macOS is over 100 MiB and RSS there is not
+ * cgroup memory. The earlier sparse-shape peak (790.3 MiB at heap 384) is
+ * lower.
+ * Measured by hand, not by this suite. Re-measure before changing it.
*/
-const D5_CAPPED_PEAK_RSS_MIB = 1129.2
+const CAPPED_PEAK_RSS_MIB = 1459.5
const MEMORY_HEADROOM = 1.25
/**
- * Owner decision on review M4 (#739): the request stays at 768Mi. It sits
- * below the limit, so the scheduler reserves what the ordinary load uses and
- * the worst-structure burst is covered by the limit alone.
+ * The request stays at 768Mi (owner decision on review M4, #739). It sits
+ * below the limit, so the scheduler reserves 768Mi while the pod can burst to
+ * the limit. That burst is not reserved on the node.
*/
const MEMORY_REQUEST_MIB = 768
@@ -96,7 +96,7 @@ describe('codex-llm-proxy base manifest', () => {
expect(grace).toBe(Math.ceil(worstCaseMs / 1000) + SHUTDOWN_MARGIN_SECONDS)
})
- it('T-DEP-2 caps the heap below the memory limit, keeps the limit above the D5 peak and requests 768Mi', () => {
+ it('T-DEP-2 caps the heap below the memory limit, keeps the limit 25% above the worst-structure peak and requests 768Mi', () => {
const nodeOptions = activeLines(MANIFEST).findIndex((line) => /name:\s*NODE_OPTIONS\s*$/.test(line))
expect(nodeOptions, 'the container must set NODE_OPTIONS').toBeGreaterThanOrEqual(0)
const heap = /--max-old-space-size=(\d+)/.exec(activeLines(MANIFEST)[nodeOptions + 1] ?? '')
@@ -106,7 +106,7 @@ describe('codex-llm-proxy base manifest', () => {
expect(Number(heap![1])).toBeLessThan(limit)
expect(resourceMemory('requests')).toBeLessThanOrEqual(limit)
expect(resourceMemory('requests')).toBe(MEMORY_REQUEST_MIB)
- expect(limit).toBeGreaterThanOrEqual(Math.ceil(D5_CAPPED_PEAK_RSS_MIB * MEMORY_HEADROOM))
+ expect(limit).toBeGreaterThanOrEqual(Math.ceil(CAPPED_PEAK_RSS_MIB * MEMORY_HEADROOM))
})
it('T-DEP-3 keeps the Codex subscription off in base, as keyper-labs/evenfire-infra CI requires', () => {
diff --git a/grok-llm-proxy/test/deployManifest.test.ts b/grok-llm-proxy/test/deployManifest.test.ts
index 3e2277025..7a55007dd 100644
--- a/grok-llm-proxy/test/deployManifest.test.ts
+++ b/grok-llm-proxy/test/deployManifest.test.ts
@@ -24,31 +24,31 @@ const WORKFLOW_RECIPES = readFileSync(
/** Headroom between the last in-flight request finishing and SIGKILL. */
const SHUTDOWN_MARGIN_SECONDS = 20
/**
- * Heap cap the peak below was measured with. At 384 MiB the process aborts
- * while it parses the three admitted ordinary bodies when each one carries the
- * worst structure the contract admits (262144 containers, 262144 members,
- * 1048576 elements), before any stream reaches the upstream. 512 MiB is the
- * smallest cap that survives that load.
+ * Heap cap the peak below was measured with. The byte-based BodyBudget admits
+ * about eight compact bodies of the worst structure the contract admits
+ * (262144 containers, 262144 members, 1048576 elements; 2.8 MiB each), and each
+ * parsed tree costs about 44 MiB after a collection. With 8 held streams, 8
+ * queued bodies and one V2 stream, a 640 MiB cap aborted the Codex proxy in 3
+ * of 3 runs (Grok passed 3 of 3); 768 MiB passed both proxies in 3 of 3.
+ * 768 MiB is the smallest cap tested that passes both, not a proven minimum.
*/
-const HEAP_CAP_MIB = 512
+const HEAP_CAP_MIB = 768
/**
- * The largest measured peak, heap capped at HEAP_CAP_MIB (tsc build, one
- * process, upstream request through undici, macOS, Node v24.18.0):
- * - D5 (#739): eight 8 MiB streams and three queued 8 MiB bodies with a
- * text tool description: 511 MiB at a 384 MiB cap;
- * - D5 plus the visual slot (#784), same text load: 775.4 MiB at 384;
- * - the same load with every ordinary body carrying the worst structure the
- * contract admits (#806 Q1), plus one ~36 MB V2 stream: 1096.8 and 1110.7
- * MiB in two runs at 512. That is the recorded peak.
- * Measured by hand, not by this suite (harness: 8 held streams, 3 queued
- * bodies, 1 visual body). Re-measure before changing it.
+ * The largest measured peak RSS, heap capped at HEAP_CAP_MIB (real proxy built
+ * with tsc, upstream request through undici, macOS, Node v24.18.0, the compact
+ * worst-structure shape: 8 held streams, 8 queued bodies, one ~36 MB V2
+ * stream). Three runs: 1357.1, 1481.6 and 1212.5 MiB. The recorded peak is the
+ * maximum, because the run-to-run spread on macOS is over 100 MiB and RSS
+ * there is not cgroup memory. The earlier sparse-shape peaks (511 MiB at
+ * heap 384, 775.4 MiB with the visual slot) are lower.
+ * Measured by hand, not by this suite. Re-measure before changing it.
*/
-const CAPPED_PEAK_RSS_MIB = 1110.7
+const CAPPED_PEAK_RSS_MIB = 1481.6
const MEMORY_HEADROOM = 1.25
/**
- * Owner decision on review M4 (#739): the request stays at 768Mi. It sits
- * below the limit, so the scheduler reserves what the ordinary load uses and
- * the worst-structure burst is covered by the limit alone.
+ * The request stays at 768Mi (owner decision on review M4, #739). It sits
+ * below the limit, so the scheduler reserves 768Mi while the pod can burst to
+ * the limit. That burst is not reserved on the node.
*/
const MEMORY_REQUEST_MIB = 768
@@ -101,7 +101,7 @@ describe('grok-llm-proxy base manifest', () => {
expect(grace).toBe(Math.ceil(worstCaseMs / 1000) + SHUTDOWN_MARGIN_SECONDS)
})
- it('T-DEP-2 caps the heap below the memory limit, keeps the limit 25% above the worst-structure-plus-visual peak and requests 768Mi', () => {
+ it('T-DEP-2 caps the heap below the memory limit, keeps the limit 25% above the worst-structure peak and requests 768Mi', () => {
const nodeOptions = activeLines(MANIFEST).findIndex((line) => /name:\s*NODE_OPTIONS\s*$/.test(line))
expect(nodeOptions, 'the container must set NODE_OPTIONS').toBeGreaterThanOrEqual(0)
const heap = /--max-old-space-size=(\d+)/.exec(activeLines(MANIFEST)[nodeOptions + 1] ?? '')
diff --git a/scripts/tests/test-codex-llm-proxy-deploy-contract.sh b/scripts/tests/test-codex-llm-proxy-deploy-contract.sh
index 1ac340c54..8ae48829e 100755
--- a/scripts/tests/test-codex-llm-proxy-deploy-contract.sh
+++ b/scripts/tests/test-codex-llm-proxy-deploy-contract.sh
@@ -99,24 +99,25 @@ if named_containers != ["codex-llm-proxy"]:
# Measured on the shipping design (#739 D5): eight 8 MiB streams, two 24 MiB
# visual streams and three queued 8 MiB bodies peaked at 790 MiB of RSS with a
-# 384 MiB old space and at 886-1009 MiB without it. When the three ordinary
-# bodies carry the worst structure the contract admits (#806 Q1), a 384 MiB old
-# space aborts while parsing them, and a 512 MiB one is the smallest that
-# survives; that load peaked at 1095-1129 MiB of RSS. The limit is that peak
-# plus 25 %, rounded up to 1536Mi. The request stays at 768Mi (owner decision on
-# review M4), so the scheduler reserves what the ordinary load uses and the
-# worst-structure burst is covered by the limit alone.
+# 384 MiB old space and at 886-1009 MiB without it. The byte-based BodyBudget
+# also admits about eight compact bodies of the worst structure the contract
+# admits (#806 Q1), each a ~44 MiB tree once parsed. With 8 held streams, 8
+# queued bodies and two visual streams, a 640 MiB old space aborted 3 of 3 and
+# 768 MiB passed 3 of 3; that load peaked at 1327-1460 MiB of RSS on macOS (not
+# cgroup memory). The limit is the maximum plus 25 %, rounded up to 2048Mi. The
+# request stays at 768Mi (owner decision on review M4), so the scheduler
+# reserves 768Mi and the burst up to the limit is not reserved on the node.
memory_limit = re.search(r"limits:\n\s+cpu: \S+\n\s+memory: (\S+)", text)
memory_request = re.search(r"requests:\n\s+cpu: \S+\n\s+memory: (\S+)", text)
heap_cap = re.search(
r"- name: NODE_OPTIONS\n\s+value: \"--max-old-space-size=(\d+)\"", text
)
-if not memory_limit or memory_limit.group(1) != "1536Mi":
- errors.append("proxy memory limit must be 1536Mi")
+if not memory_limit or memory_limit.group(1) != "2048Mi":
+ errors.append("proxy memory limit must be 2048Mi")
if not memory_request or memory_request.group(1) != "768Mi":
errors.append("proxy memory request must be 768Mi")
-if not heap_cap or heap_cap.group(1) != "512":
- errors.append("proxy must cap the V8 old space at 512 MiB through NODE_OPTIONS")
+if not heap_cap or heap_cap.group(1) != "768":
+ errors.append("proxy must cap the V8 old space at 768 MiB through NODE_OPTIONS")
# #739 D6: on SIGTERM the proxy stops accepting and waits for its open streams
# (main.ts awaits servers.close()), so the grace period must outlast the
diff --git a/scripts/tests/test-grok-llm-proxy-deploy-contract.sh b/scripts/tests/test-grok-llm-proxy-deploy-contract.sh
index 099114419..1f372e3d6 100755
--- a/scripts/tests/test-grok-llm-proxy-deploy-contract.sh
+++ b/scripts/tests/test-grok-llm-proxy-deploy-contract.sh
@@ -87,24 +87,26 @@ if "CONTROL_API_INTERNAL_SERVICE_TOKENS" in text:
# Eight 8 MiB streams and three queued 8 MiB bodies (#739 D5) peaked at 511 MiB
# of RSS with a 384 MiB old space (#739 measured 509.8); one ~36 MB V2 stream in
-# the visual slot (#784) raised that to 775 MiB. When the three ordinary bodies
-# carry the worst structure the contract admits (#806 Q1), a 384 MiB old space
-# aborts while parsing them, and a 512 MiB one is the smallest that survives;
-# with one visual stream that load peaked at 1097-1111 MiB of RSS. The limit is
-# that peak plus 25 %, rounded up to 1536Mi. The request stays at 768Mi (owner
-# decision on review M4), so the scheduler reserves what the ordinary load uses
-# and the worst-structure burst is covered by the limit alone.
+# the visual slot (#784) raised that to 775 MiB. The byte-based BodyBudget also
+# admits about eight compact bodies of the worst structure the contract admits
+# (#806 Q1), each a ~44 MiB tree once parsed. With 8 held streams, 8 queued
+# bodies and one V2 stream, a 640 MiB old space passed 3 of 3 for Grok and
+# aborted 3 of 3 for Codex, and 768 MiB passed both in 3 of 3; that load peaked
+# at 1212-1482 MiB of RSS on macOS (not cgroup memory). The limit is the
+# maximum plus 25 %, rounded up to 2048Mi. The request stays at 768Mi (owner
+# decision on review M4), so the scheduler reserves 768Mi and the burst up to
+# the limit is not reserved on the node.
memory_limit = re.search(r"limits:\n\s+cpu: \S+\n\s+memory: (\S+)", text)
memory_request = re.search(r"requests:\n\s+cpu: \S+\n\s+memory: (\S+)", text)
heap_cap = re.search(
r"- name: NODE_OPTIONS\n\s+value: \"--max-old-space-size=(\d+)\"", text
)
-if not memory_limit or memory_limit.group(1) != "1536Mi":
- errors.append("proxy memory limit must be 1536Mi")
+if not memory_limit or memory_limit.group(1) != "2048Mi":
+ errors.append("proxy memory limit must be 2048Mi")
if not memory_request or memory_request.group(1) != "768Mi":
errors.append("proxy memory request must be 768Mi")
-if not heap_cap or heap_cap.group(1) != "512":
- errors.append("proxy must cap the V8 old space at 512 MiB through NODE_OPTIONS")
+if not heap_cap or heap_cap.group(1) != "768":
+ errors.append("proxy must cap the V8 old space at 768 MiB through NODE_OPTIONS")
# #739 D6: on SIGTERM the proxy stops accepting and waits for its open streams
# (main.ts awaits servers.close()), so the grace period must outlast the