feat(orch): per-start UFFD startup working-set metric - #2960
Conversation
Add per-handler cumulative counters for demand faults served — pages resolved, the subset pulled from the source, and bytes installed — and a ServeStats() snapshot to read them without blocking. Surface it through the MemoryBackend interface, with a passthrough in Uffd and a zero stub in NoopMemory. These mirror the orchestrator.sandbox.uffd.serve metric but as a point-in-time snapshot, so a caller can later sample how many pages a guest needed to start. No behaviour change. Signed-off-by: Nikita Kalyazin <nikita.kalyazin@e2b.dev>
There is no fleet-wide signal for how many pages or how much data a guest
needs to start (until envd init returns). The uffd.serve metric counts
demand faults over the whole sandbox lifetime and has no per-sandbox
dimension, so it can give a fleet rate but not a per-start distribution.
Record orchestrator.sandbox.uffd.startup.{pages,source_pages,bytes} once
per start, at the envd-init boundary, with the start's serve-stats
snapshot as the sample value. Because the observation unit is one start,
histogram_quantile yields the per-sandbox avg/p50/p95. Attributes
start_type (create/resume) and success let slow or failed starts be
correlated with page volume. source_pages isolates the data actually
pulled from the source (e.g. GCS) and doubles as a prefetch-hit signal.
Signed-off-by: Nikita Kalyazin <nikita.kalyazin@e2b.dev>
Assert ServeStats() folds a mix of finished serve attempts correctly: resolved faults (installed/present) count as needed pages, the source subset and installed bytes are tracked, and deferred/errored attempts are excluded so a deferred fault is not double-counted when re-served. Signed-off-by: Nikita Kalyazin <nikita.kalyazin@e2b.dev>
PR SummaryLow Risk Overview The UFFD handler now keeps cumulative ServeStats (resolved pages, source-backed pages, installed bytes) updated on each finished serve; Reviewed by Cursor Bugbot for commit 0374cbb. Bugbot is set up for automated code reviews on this repo. Configure here. |
There was a problem hiding this comment.
Code Review
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
❌ 5 Tests Failed:
View the top 2 failed test(s) by shortest run time
View the full list of 3 ❄️ flaky test(s)
To view more test analytics, go to the Test Analytics Dashboard |
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 51dd901. Configure here.
ServeStats() is lifetime-cumulative on the UFFD handler, and WaitForEnvd can run more than once on the same handler: a template build swaps the envd binary and restarts the service mid-build, then waits again. That second wait would emit an orchestrator.sandbox.uffd.startup.* sample holding the cumulative serve counts — startup pages plus all the build-step demand faults since — rather than that init's working set, polluting the distribution. Guard the recording with a sync.Once so it fires only on the first wait, which is the actual sandbox start and the point at which the cumulative counters still equal the startup counts. Signed-off-by: Nikita Kalyazin <nikita.kalyazin@e2b.dev>
## Why
No fleet-wide signal exists for **how many pages / how much data a
sandbox needs
to start** (until envd init returns). `uffd.serve` can't answer it: it's
lifetime-cumulative (not bounded to startup) and has no per-sandbox
dimension
(so Mimir gives a fleet rate, never a per-sandbox p50/p95).
## What
A metric recorded **once per start** at the envd-init boundary, with
that
start's serve-stats snapshot as the value. One observation per start in
a native
histogram → `histogram_quantile` gives the **across-sandbox** p50/p95
directly.
| Metric | Meaning |
|---|---|
| `orchestrator.sandbox.uffd.startup.pages` | Demand-fault pages
resolved to reach envd init — the headline "pages to start". |
| `orchestrator.sandbox.uffd.startup.source_pages` | Subset pulled from
the source (GCS/NFS) — the pages that cost a real read. The rest were
zero-filled or already resident (prefetch hits). |
| `orchestrator.sandbox.uffd.startup.bytes` | Bytes faulted into the
guest — the "how much data" answer. |
Attributes: `start_type` ∈ `{create, resume}`, `success` ∈ `{true,
false}`. No
high-cardinality labels. Counts demand faults only (resolved, no
double-counting); prefaulted pages bypass the serve loop
---------
Signed-off-by: Nikita Kalyazin <nikita.kalyazin@e2b.dev>
🤖 I have created a release *beep* *boop* --- ## 0.0.1 (2026-07-30) ### Features * **api:** add sandbox IAM workload token configuration ([13ddb3d](13ddb3d)) * **api:** add sandbox workload identity permission ([#3319](#3319)) ([13ddb3d](13ddb3d)) * **api:** SOCKS5 egress proxy on sandbox network config (BYOP) ([#2642](#2642)) ([1fc3820](1fc3820)) * **cfg:** add DISABLE_STARTUP_RECLAIM flag ([#3081](#3081)) ([7677ca6](7677ca6)) * **clickhouse:** implement multi-cluster fan-out for events and stats ([#2925](#2925)) ([39594c6](39594c6)) * dynamic sandbox log routing and ClickHouse-backed log reads ([#3236](#3236)) ([1b19a3b](1b19a3b)) * **envd:** give envd realtime IO priority, reset for user processes ([#2681](#2681)) ([f4bd1b2](f4bd1b2)) * **envd:** split collapse stats into real migrations vs already-huge ([#3021](#3021)) ([0d77614](0d77614)) * **envd:** support user-defined file metadata via xattrs ([#2732](#2732)) ([da8fbe4](da8fbe4)) * **featureflags:** support per-service context providers ([#3100](#3100)) ([65297c1](65297c1)) * freeze user cgroup across pause/resume to keep envd /init responsive ([#2688](#2688)) ([eceb741](eceb741)) * **metrics:** break down pause-snapshot latency by step ([#3426](#3426)) ([657559e](657559e)) * **metrics:** label pause telemetry by fs_only ([#3425](#3425)) ([411b63e](411b63e)) * **observability:** add kill_reason to sandbox.lifecycle.killed ([#2833](#2833)) ([e45418f](e45418f)) * **observability:** include kill_reason in kill-path structured logs ([#2846](#2846)) ([33c49f7](33c49f7)) * **orch:** add envd-version to LaunchDarkly sandbox context ([#3051](#3051)) ([37d3b92](37d3b92)) * **orch:** add less, nftables, iputils-ping, and jq to base provisioning ([#2736](#2736)) ([a1e010e](a1e010e)) * **orch:** collapse envd's heap into 2 MiB hugepages before pause to cut cold-resume faults ([#2997](#2997)) ([6677f73](6677f73)) * **orch:** debug a sandbox guest kernel with resume-build -gdb ([#3040](#3040)) ([37bb0dc](37bb0dc)) * **orch:** decouple warm resume from memfile dedup ([#3166](#3166)) ([77f25a0](77f25a0)) * **orch:** distro-aware template base-image provisioning ([#3411](#3411)) ([f8c7b5b](f8c7b5b)) * **orchestrator/cgroup:** list and destroy leaked sandbox cgroups ([#3086](#3086)) ([bce1d84](bce1d84)) * **orchestrator/nbd:** inspect and disconnect connected devices ([#3087](#3087)) ([4d47148](4d47148)) * **orchestrator/network:** list slot namespaces ([#3089](#3089)) ([c23dbc7](c23dbc7)) * **orchestrator/network:** list slot namespaces ([#3090](#3090)) ([fbfce25](fbfce25)) * **orchestrator:** add -force-reboot to resume-build to cold-boot memory-snaphsot builds ([#3208](#3208)) ([cf8f15b](cf8f15b)) * **orchestrator:** add allocated resource metrics for sandboxes ([#2943](#2943)) ([95cb6d3](95cb6d3)) * **orchestrator:** add dummy orchestrator binary for local API dev ([#2744](#2744)) ([ab56e25](ab56e25)) * **orchestrator:** add NetworkAssignHook for sandbox lifecycle extensions ([#3290](#3290)) ([3261963](3261963)) * **orchestrator:** add soft-delete marker label to the check metric ([#3144](#3144)) ([1ce64f8](1ce64f8)) * **orchestrator:** add v4HeaderForUncompressed FF bit ([#2669](#2669)) ([1f459ee](1f459ee)) * **orchestrator:** always include execution metrics in sandbox webhook events ([#2852](#2852)) ([440edfe](440edfe)) * **orchestrator:** classify envd-init by exit type ([#3139](#3139)) ([1e39a4f](1e39a4f)) * **orchestrator:** graceful sandbox drain on shutdown ([#3069](#3069)) ([6ce68e3](6ce68e3)) * **orchestrator:** graceful template-build drain on shutdown ([#3079](#3079)) ([1b3001c](1b3001c)) * **orchestrator:** improved read-path telemetry ([#3063](#3063)) ([bc3fe84](bc3fe84)) * **orchestrator:** LD-gated ClickHouse write fan-out feature flag ([#3152](#3152)) ([f046fcf](f046fcf)) * **orchestrator:** make build-reserved-disk-space-mb default 256MB ([#3065](#3065)) ([d473f98](d473f98)) * **orchestrator:** record upload compression metrics ([#2761](#2761)) ([9092e35](9092e35)) * **orchestrator:** report hugepage metrics to API ([#3182](#3182)) ([7735bae](7735bae)) * **orchestrator:** run startup reclaim on boot ([#3123](#3123)) ([79b838e](79b838e)) * **orchestrator:** single-instance flock on startup ([#3143](#3143)) ([1320d6e](1320d6e)) * **orchestrator:** soft-delete consumer enforcement for storage index ([#3034](#3034)) ([fbfc918](fbfc918)) * **orchestrator:** tag envd-init meters with start_type ([#3125](#3125)) ([4466b48](4466b48)) * **orchestrator:** track and report last status change timestamp ([#2980](#2980)) ([f79be77](f79be77)) * **orchestrator:** track sandbox lifecycles ([#2998](#2998)) ([057f20c](057f20c)) * **orchestrator:** write layer sizes (logical/mapped/diff) to object metadata ([#3122](#3122)) ([11869c0](11869c0)) * **orch:** harvest resume-prefetch trace on pause ([#3067](#3067)) ([97bd4a5](97bd4a5)) * **orch:** last-cycle memory prefetch on resume ([#3258](#3258)) ([b22e820](b22e820)) * **orch:** make resume-build -gdb work on real nodes + add copy-build -gdb ([#3108](#3108)) ([c684bd2](c684bd2)) * **orch:** opt-in DSCP marker for sandbox egress (SANDBOX_EGRESS_DSCP) ([#3039](#3039)) ([a98cf2c](a98cf2c)) * **orch:** per-start UFFD startup working-set metric ([#2960](#2960)) ([dc386b2](dc386b2)) * **orch:** premade NixOS base-image support ([#3412](#3412)) ([776ba39](776ba39)) * **orch:** record envd init duration histogram on failure with success attribute ([#2749](#2749)) ([afa7458](afa7458)) * **orch:** snapshot fragmentation metrics ([#2931](#2931)) ([842b007](842b007)) * per-team events TTL limit (tier + addons) ([#3181](#3181)) ([f76b2cb](f76b2cb)) * **shared:** add OTEL instrumentation to AWS S3 storage client ([#3172](#3172)) ([25b0fd1](25b0fd1)) * **storage:** per-role storage URLs, env-free storage library ([#3246](#3246)) ([fcbe909](fcbe909)) * **storage:** stamp provenance custom metadata on uploaded objects (incl. headers) ([#3033](#3033)) ([ba8604e](ba8604e)) * **storage:** write-through compressed templates to NFS on upload ([#2827](#2827)) ([57503c1](57503c1)) ### Bug Fixes * added api and orch ([#3454](#3454)) ([fda5e45](fda5e45)) * **block:** rephrase misleading error message in pwritevAll ([#2816](#2816)) ([1555f1b](1555f1b)) * **cache:** use 512-byte units for stat.Blocks in FileSize ([#2949](#2949)) ([0f632a9](0f632a9)) * **clean-nfs-cache:** exclude zombies from delete_age ([#3191](#3191)) ([3fa2aeb](3fa2aeb)) * **compression:** correctness findings from compression audit ([#2803](#2803)) ([d21a6a9](d21a6a9)) * **copy-build:** resolve compression suffix for build data files ([#2859](#2859)) ([8966f7e](8966f7e)) * correct 3 CVES ([#3218](#3218)) ([076823b](076823b)) * **envd:** stop freezing socat cgroup across pause/resume ([#2923](#2923)) ([8b6f2b9](8b6f2b9)) * **inspect-build:** adapt validate to new Chunker upstream API ([#2989](#2989)) ([2e0d3da](2e0d3da)) * **nbd:** adjust status poll sleep from 100ns to 100µs ([02bf51b](02bf51b)) * **nbd:** change NBD status poll sleep from 100ns to 100µs to avoid useless busy spinning ([#2884](#2884)) ([02bf51b](02bf51b)) * **nfsproxy:** deflake TestRoundTrip EADDRINUSE ([#2987](#2987)) ([55f4d18](55f4d18)) * **orch:** denormalize upload metric file type ([#2865](#2865)) ([b1646ca](b1646ca)) * **orch:** disable the chronyd seccomp filter on Alpine when using PHC ([#3453](#3453)) ([e58af28](e58af28)) * **orchestrator:** anchor rsync CWD to root in template file copy ([#2835](#2835)) ([7160db9](7160db9)) * **orchestrator:** atomically replace metadata ([#3321](#3321)) ([0c4ad6b](0c4ad6b)) * **orchestrator:** avoid serializing upload headers twice ([#2762](#2762)) ([9b7b149](9b7b149)) * **orchestrator:** chunk readiness bug in P2P->compressed ([#3185](#3185)) ([74a6e5b](74a6e5b)) * **orchestrator:** deschedule flaky eviction-loop race in TestDiffSto… ([#3173](#3173)) ([88ff17c](88ff17c)) * **orchestrator:** discard poisoned nftables conn on firewall errors ([#3008](#3008)) ([03f10e0](03f10e0)) * **orchestrator:** drop stale pre-init logs ([#3297](#3297)) ([8ec4be5](8ec4be5)) * **orchestrator:** emit compression ratios as fractions, not BP ([#2772](#2772)) ([866f4c1](866f4c1)) * **orchestrator:** export dirty-page stall counter from process start ([#2992](#2992)) ([badc8ad](badc8ad)) * **orchestrator:** harden Firecracker process shutdown ([#2996](#2996)) ([df662e7](df662e7)) * **orchestrator:** harden shutdown network cleanup ([#3000](#3000)) ([de2f391](de2f391)) * **orchestrator:** implement Docker COPY merge semantics in template builds ([#3283](#3283)) ([9174104](9174104)) * **orchestrator:** keep dedup empty-pages telemetry scan-only ([#2991](#2991)) ([35d0832](35d0832)) * **orchestrator:** let build-cache threshold flag raise above its fal… ([#3175](#3175)) ([06393c3](06393c3)) * **orchestrator:** log missing egress proxy in startup reclaim instead of defaulting silently ([#3116](#3116)) ([6ca3163](6ca3163)) * **orchestrator:** make copy-build handle filesystem-only snapshots ([#3299](#3299)) ([62add04](62add04)) * **orchestrator:** measure ext4 free space from block groups ([#3282](#3282)) ([f18f05f](f18f05f)) * **orchestrator:** normalize upload metric file labels ([#2767](#2767)) ([6dec8b3](6dec8b3)) * **orchestrator:** order egress config/firewall updates to close BYOP enable race ([#3313](#3313)) ([7faa59e](7faa59e)) * **orchestrator:** order envd.service after local-fs.target ([#3043](#3043)) ([ea2663e](ea2663e)) * **orchestrator:** order envd.service after systemd-tmpfiles-setup ([#3130](#3130)) ([9481811](9481811)) * **orchestrator:** pause upload retain retry ([#2993](#2993)) ([4f81799](4f81799)) * **orchestrator:** pin tap device host-side MAC address ([#3271](#3271)) ([3c786ba](3c786ba)) * **orchestrator:** pin UFFD copy source buffers ([#2745](#2745)) ([837fa91](837fa91)) * **orchestrator:** preserve full ENV value across stdout chunks ([#2740](#2740)) ([4822e6d](4822e6d)) * **orchestrator:** read V3 ancestors as uncompressed instead of failing ([#2994](#2994)) ([c479dd3](c479dd3)) * **orchestrator:** reject standby while draining ([#3325](#3325)) ([475a7ee](475a7ee)) * **orchestrator:** report real V4 header compression ratio ([#2771](#2771)) ([ecd344e](ecd344e)) * **orchestrator:** resolve remaining P2P/compression/V5 issues ([#3015](#3015)) ([1e4379e](1e4379e)) * **orchestrator:** sanitize OCI pull errors ([#3096](#3096)) ([a3af6c0](a3af6c0)) * **orchestrator:** scope rootfs hash to provision default ([#3129](#3129)) ([475f955](475f955)) * **orchestrator:** stop Checks health-loop leaking ([#2739](#2739)) ([17e6e60](17e6e60)) * **orchestrator:** survive SIGBUS from failing disks under mmap'd caches ([#3385](#3385)) ([728bba3](728bba3)) * **orchestrator:** tolerate missing header for legacy templates ([#3026](#3026)) ([8a44bfe](8a44bfe)) * **orch:** fall back to ID_LIKE with a warning instead of rejecting ([#3459](#3459)) ([7167818](7167818)) * **orch:** prevent NBD dispatch read-loop stall on WRITE_ZEROES (behind flag) ([#3048](#3048)) ([efd3d4d](efd3d4d)) * **orch:** split scheduling base build id per artifact ([#2920](#2920)) ([3e35a2a](3e35a2a)) * **orch:** validate copy-build -gdb buckets before the snapshot copy ([#3446](#3446)) ([586ad74](586ad74)) * **shared:** never report a failed envd command stream as success ([#3281](#3281)) ([69c06b6](69c06b6)) * **storage:** compression upload & cache correctness fixes ([#3231](#3231)) ([980748f](980748f)) * **storage:** don't assume V4+ ancestor gaps are uncompressed ([#3447](#3447)) ([bfdbb24](bfdbb24)) * **uffd:** dedupe deferred page faults ([#2864](#2864)) ([9680a41](9680a41)) * WrapContextAsUserError should not misclassify internal timeouts as user cancellations ([#3155](#3155)) ([8f83959](8f83959)) ### Performance Improvements * **build:** cache resolved Diff per BuildId within File.ReadAt ([#2838](#2838)) ([53de07f](53de07f)) * **build:** parallelize fragmented backing reads ([#2872](#2872)) ([c7655a7](c7655a7)) * **clean-nfs-cache:** restore dirfd-relative statx ([#2766](#2766)) ([6bdbedb](6bdbedb)) * **header:** add V5 columnar varint header format ([#2847](#2847)) ([9dd931b](9dd931b)) * **header:** pack cached Header.Mapping into a compact form ([#2844](#2844)) ([7f0b13c](7f0b13c)) * **orchestrator:** add memfile dedup density threshold ([#2862](#2862)) ([7ccfa02](7ccfa02)) * **orchestrator:** avoid V3-ancestor header refresh ([#2999](#2999)) ([cb6aa0b](cb6aa0b)) * **orch:** metrics for dirty page throttling ([#2858](#2858)) ([d2aa554](d2aa554)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). Co-authored-by: e2b-release-please[bot] <298072688+e2b-release-please[bot]@users.noreply.github.com> Co-authored-by: Charlie Wyse <charlie.wyse@e2b.dev>
🤖 I have created a release *beep* *boop* --- ## 0.0.1 (2026-07-30) ### Features * **api:** add sandbox IAM workload token configuration ([13ddb3d](13ddb3d)) * **api:** add sandbox workload identity permission ([#3319](#3319)) ([13ddb3d](13ddb3d)) * **api:** SOCKS5 egress proxy on sandbox network config (BYOP) ([#2642](#2642)) ([1fc3820](1fc3820)) * **cfg:** add DISABLE_STARTUP_RECLAIM flag ([#3081](#3081)) ([7677ca6](7677ca6)) * **clickhouse:** implement multi-cluster fan-out for events and stats ([#2925](#2925)) ([39594c6](39594c6)) * dynamic sandbox log routing and ClickHouse-backed log reads ([#3236](#3236)) ([1b19a3b](1b19a3b)) * **envd:** give envd realtime IO priority, reset for user processes ([#2681](#2681)) ([f4bd1b2](f4bd1b2)) * **envd:** split collapse stats into real migrations vs already-huge ([#3021](#3021)) ([0d77614](0d77614)) * **envd:** support user-defined file metadata via xattrs ([#2732](#2732)) ([da8fbe4](da8fbe4)) * **featureflags:** support per-service context providers ([#3100](#3100)) ([65297c1](65297c1)) * freeze user cgroup across pause/resume to keep envd /init responsive ([#2688](#2688)) ([eceb741](eceb741)) * **metrics:** break down pause-snapshot latency by step ([#3426](#3426)) ([f551118](f551118)) * **metrics:** label pause telemetry by fs_only ([#3425](#3425)) ([4be33ba](4be33ba)) * **observability:** add kill_reason to sandbox.lifecycle.killed ([#2833](#2833)) ([e45418f](e45418f)) * **observability:** include kill_reason in kill-path structured logs ([#2846](#2846)) ([33c49f7](33c49f7)) * **orch:** add envd-version to LaunchDarkly sandbox context ([#3051](#3051)) ([37d3b92](37d3b92)) * **orch:** add less, nftables, iputils-ping, and jq to base provisioning ([#2736](#2736)) ([a1e010e](a1e010e)) * **orch:** collapse envd's heap into 2 MiB hugepages before pause to cut cold-resume faults ([#2997](#2997)) ([6677f73](6677f73)) * **orch:** debug a sandbox guest kernel with resume-build -gdb ([#3040](#3040)) ([37bb0dc](37bb0dc)) * **orch:** decouple warm resume from memfile dedup ([#3166](#3166)) ([77f25a0](77f25a0)) * **orch:** distro-aware template base-image provisioning ([#3411](#3411)) ([1abece1](1abece1)) * **orchestrator/cgroup:** list and destroy leaked sandbox cgroups ([#3086](#3086)) ([bce1d84](bce1d84)) * **orchestrator/nbd:** inspect and disconnect connected devices ([#3087](#3087)) ([4d47148](4d47148)) * **orchestrator/network:** list slot namespaces ([#3089](#3089)) ([c23dbc7](c23dbc7)) * **orchestrator/network:** list slot namespaces ([#3090](#3090)) ([fbfce25](fbfce25)) * **orchestrator:** add -force-reboot to resume-build to cold-boot memory-snaphsot builds ([#3208](#3208)) ([cf8f15b](cf8f15b)) * **orchestrator:** add allocated resource metrics for sandboxes ([#2943](#2943)) ([95cb6d3](95cb6d3)) * **orchestrator:** add dummy orchestrator binary for local API dev ([#2744](#2744)) ([ab56e25](ab56e25)) * **orchestrator:** add NetworkAssignHook for sandbox lifecycle extensions ([#3290](#3290)) ([3261963](3261963)) * **orchestrator:** add soft-delete marker label to the check metric ([#3144](#3144)) ([1ce64f8](1ce64f8)) * **orchestrator:** add v4HeaderForUncompressed FF bit ([#2669](#2669)) ([1f459ee](1f459ee)) * **orchestrator:** always include execution metrics in sandbox webhook events ([#2852](#2852)) ([440edfe](440edfe)) * **orchestrator:** classify envd-init by exit type ([#3139](#3139)) ([1e39a4f](1e39a4f)) * **orchestrator:** graceful sandbox drain on shutdown ([#3069](#3069)) ([6ce68e3](6ce68e3)) * **orchestrator:** graceful template-build drain on shutdown ([#3079](#3079)) ([1b3001c](1b3001c)) * **orchestrator:** improved read-path telemetry ([#3063](#3063)) ([bc3fe84](bc3fe84)) * **orchestrator:** LD-gated ClickHouse write fan-out feature flag ([#3152](#3152)) ([f046fcf](f046fcf)) * **orchestrator:** make build-reserved-disk-space-mb default 256MB ([#3065](#3065)) ([d473f98](d473f98)) * **orchestrator:** record upload compression metrics ([#2761](#2761)) ([9092e35](9092e35)) * **orchestrator:** report hugepage metrics to API ([#3182](#3182)) ([7735bae](7735bae)) * **orchestrator:** run startup reclaim on boot ([#3123](#3123)) ([79b838e](79b838e)) * **orchestrator:** single-instance flock on startup ([#3143](#3143)) ([1320d6e](1320d6e)) * **orchestrator:** soft-delete consumer enforcement for storage index ([#3034](#3034)) ([fbfc918](fbfc918)) * **orchestrator:** tag envd-init meters with start_type ([#3125](#3125)) ([4466b48](4466b48)) * **orchestrator:** track and report last status change timestamp ([#2980](#2980)) ([f79be77](f79be77)) * **orchestrator:** track sandbox lifecycles ([#2998](#2998)) ([057f20c](057f20c)) * **orchestrator:** write layer sizes (logical/mapped/diff) to object metadata ([#3122](#3122)) ([11869c0](11869c0)) * **orch:** harvest resume-prefetch trace on pause ([#3067](#3067)) ([97bd4a5](97bd4a5)) * **orch:** last-cycle memory prefetch on resume ([#3258](#3258)) ([ea94196](ea94196)) * **orch:** make resume-build -gdb work on real nodes + add copy-build -gdb ([#3108](#3108)) ([5385594](5385594)) * **orch:** opt-in DSCP marker for sandbox egress (SANDBOX_EGRESS_DSCP) ([#3039](#3039)) ([a98cf2c](a98cf2c)) * **orch:** per-start UFFD startup working-set metric ([#2960](#2960)) ([dc386b2](dc386b2)) * **orch:** premade NixOS base-image support ([#3412](#3412)) ([4bd42d2](4bd42d2)) * **orch:** record envd init duration histogram on failure with success attribute ([#2749](#2749)) ([afa7458](afa7458)) * **orch:** snapshot fragmentation metrics ([#2931](#2931)) ([842b007](842b007)) * per-team events TTL limit (tier + addons) ([#3181](#3181)) ([f76b2cb](f76b2cb)) * **shared:** add OTEL instrumentation to AWS S3 storage client ([#3172](#3172)) ([25b0fd1](25b0fd1)) * **storage:** per-role storage URLs, env-free storage library ([#3246](#3246)) ([fcbe909](fcbe909)) * **storage:** stamp provenance custom metadata on uploaded objects (incl. headers) ([#3033](#3033)) ([ba8604e](ba8604e)) * **storage:** write-through compressed templates to NFS on upload ([#2827](#2827)) ([57503c1](57503c1)) ### Bug Fixes * added api and orch ([#3454](#3454)) ([d56e0a8](d56e0a8)) * **block:** rephrase misleading error message in pwritevAll ([#2816](#2816)) ([1555f1b](1555f1b)) * **cache:** use 512-byte units for stat.Blocks in FileSize ([#2949](#2949)) ([0f632a9](0f632a9)) * **clean-nfs-cache:** exclude zombies from delete_age ([#3191](#3191)) ([3fa2aeb](3fa2aeb)) * **compression:** correctness findings from compression audit ([#2803](#2803)) ([d21a6a9](d21a6a9)) * **copy-build:** resolve compression suffix for build data files ([#2859](#2859)) ([8966f7e](8966f7e)) * correct 3 CVES ([#3218](#3218)) ([076823b](076823b)) * **envd:** stop freezing socat cgroup across pause/resume ([#2923](#2923)) ([8b6f2b9](8b6f2b9)) * **inspect-build:** adapt validate to new Chunker upstream API ([#2989](#2989)) ([2e0d3da](2e0d3da)) * **nbd:** adjust status poll sleep from 100ns to 100µs ([02bf51b](02bf51b)) * **nbd:** change NBD status poll sleep from 100ns to 100µs to avoid useless busy spinning ([#2884](#2884)) ([02bf51b](02bf51b)) * **nfsproxy:** deflake TestRoundTrip EADDRINUSE ([#2987](#2987)) ([55f4d18](55f4d18)) * **orch:** denormalize upload metric file type ([#2865](#2865)) ([b1646ca](b1646ca)) * **orch:** disable the chronyd seccomp filter on Alpine when using PHC ([#3453](#3453)) ([dfa9764](dfa9764)) * **orchestrator:** anchor rsync CWD to root in template file copy ([#2835](#2835)) ([7160db9](7160db9)) * **orchestrator:** atomically replace metadata ([#3321](#3321)) ([0c4ad6b](0c4ad6b)) * **orchestrator:** avoid serializing upload headers twice ([#2762](#2762)) ([9b7b149](9b7b149)) * **orchestrator:** chunk readiness bug in P2P->compressed ([#3185](#3185)) ([74a6e5b](74a6e5b)) * **orchestrator:** deschedule flaky eviction-loop race in TestDiffSto… ([#3173](#3173)) ([88ff17c](88ff17c)) * **orchestrator:** discard poisoned nftables conn on firewall errors ([#3008](#3008)) ([03f10e0](03f10e0)) * **orchestrator:** drop stale pre-init logs ([#3297](#3297)) ([8ec4be5](8ec4be5)) * **orchestrator:** emit compression ratios as fractions, not BP ([#2772](#2772)) ([866f4c1](866f4c1)) * **orchestrator:** export dirty-page stall counter from process start ([#2992](#2992)) ([badc8ad](badc8ad)) * **orchestrator:** harden Firecracker process shutdown ([#2996](#2996)) ([df662e7](df662e7)) * **orchestrator:** harden shutdown network cleanup ([#3000](#3000)) ([de2f391](de2f391)) * **orchestrator:** implement Docker COPY merge semantics in template builds ([#3283](#3283)) ([9174104](9174104)) * **orchestrator:** keep dedup empty-pages telemetry scan-only ([#2991](#2991)) ([35d0832](35d0832)) * **orchestrator:** let build-cache threshold flag raise above its fal… ([#3175](#3175)) ([06393c3](06393c3)) * **orchestrator:** log missing egress proxy in startup reclaim instead of defaulting silently ([#3116](#3116)) ([6ca3163](6ca3163)) * **orchestrator:** make copy-build handle filesystem-only snapshots ([#3299](#3299)) ([62add04](62add04)) * **orchestrator:** measure ext4 free space from block groups ([#3282](#3282)) ([f18f05f](f18f05f)) * **orchestrator:** normalize upload metric file labels ([#2767](#2767)) ([6dec8b3](6dec8b3)) * **orchestrator:** order egress config/firewall updates to close BYOP enable race ([#3313](#3313)) ([7faa59e](7faa59e)) * **orchestrator:** order envd.service after local-fs.target ([#3043](#3043)) ([ea2663e](ea2663e)) * **orchestrator:** order envd.service after systemd-tmpfiles-setup ([#3130](#3130)) ([9481811](9481811)) * **orchestrator:** pause upload retain retry ([#2993](#2993)) ([4f81799](4f81799)) * **orchestrator:** pin tap device host-side MAC address ([#3271](#3271)) ([3c786ba](3c786ba)) * **orchestrator:** pin UFFD copy source buffers ([#2745](#2745)) ([837fa91](837fa91)) * **orchestrator:** preserve full ENV value across stdout chunks ([#2740](#2740)) ([4822e6d](4822e6d)) * **orchestrator:** read V3 ancestors as uncompressed instead of failing ([#2994](#2994)) ([c479dd3](c479dd3)) * **orchestrator:** reject standby while draining ([#3325](#3325)) ([475a7ee](475a7ee)) * **orchestrator:** report real V4 header compression ratio ([#2771](#2771)) ([ecd344e](ecd344e)) * **orchestrator:** resolve remaining P2P/compression/V5 issues ([#3015](#3015)) ([1e4379e](1e4379e)) * **orchestrator:** sanitize OCI pull errors ([#3096](#3096)) ([a3af6c0](a3af6c0)) * **orchestrator:** scope rootfs hash to provision default ([#3129](#3129)) ([475f955](475f955)) * **orchestrator:** stop Checks health-loop leaking ([#2739](#2739)) ([17e6e60](17e6e60)) * **orchestrator:** survive SIGBUS from failing disks under mmap'd caches ([#3385](#3385)) ([8694d08](8694d08)) * **orchestrator:** tolerate missing header for legacy templates ([#3026](#3026)) ([8a44bfe](8a44bfe)) * **orch:** fall back to ID_LIKE with a warning instead of rejecting ([#3459](#3459)) ([73399b3](73399b3)) * **orch:** prevent NBD dispatch read-loop stall on WRITE_ZEROES (behind flag) ([#3048](#3048)) ([efd3d4d](efd3d4d)) * **orch:** split scheduling base build id per artifact ([#2920](#2920)) ([3e35a2a](3e35a2a)) * **orch:** validate copy-build -gdb buckets before the snapshot copy ([#3446](#3446)) ([9be382f](9be382f)) * **shared:** never report a failed envd command stream as success ([#3281](#3281)) ([69c06b6](69c06b6)) * **storage:** compression upload & cache correctness fixes ([#3231](#3231)) ([980748f](980748f)) * **storage:** don't assume V4+ ancestor gaps are uncompressed ([#3447](#3447)) ([f828d12](f828d12)) * **uffd:** dedupe deferred page faults ([#2864](#2864)) ([9680a41](9680a41)) * WrapContextAsUserError should not misclassify internal timeouts as user cancellations ([#3155](#3155)) ([8f83959](8f83959)) ### Performance Improvements * **build:** cache resolved Diff per BuildId within File.ReadAt ([#2838](#2838)) ([53de07f](53de07f)) * **build:** parallelize fragmented backing reads ([#2872](#2872)) ([c7655a7](c7655a7)) * **clean-nfs-cache:** restore dirfd-relative statx ([#2766](#2766)) ([6bdbedb](6bdbedb)) * **header:** add V5 columnar varint header format ([#2847](#2847)) ([9dd931b](9dd931b)) * **header:** pack cached Header.Mapping into a compact form ([#2844](#2844)) ([7f0b13c](7f0b13c)) * **orchestrator:** add memfile dedup density threshold ([#2862](#2862)) ([7ccfa02](7ccfa02)) * **orchestrator:** avoid V3-ancestor header refresh ([#2999](#2999)) ([cb6aa0b](cb6aa0b)) * **orch:** metrics for dirty page throttling ([#2858](#2858)) ([d2aa554](d2aa554)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). Co-authored-by: e2b-release-please[bot] <298072688+e2b-release-please[bot]@users.noreply.github.com> Co-authored-by: Charlie Wyse <charlie.wyse@e2b.dev>

Why
No fleet-wide signal exists for how many pages / how much data a sandbox needs
to start (until envd init returns).
uffd.servecan't answer it: it'slifetime-cumulative (not bounded to startup) and has no per-sandbox dimension
(so Mimir gives a fleet rate, never a per-sandbox p50/p95).
What
A metric recorded once per start at the envd-init boundary, with that
start's serve-stats snapshot as the value. One observation per start in a native
histogram →
histogram_quantilegives the across-sandbox p50/p95 directly.orchestrator.sandbox.uffd.startup.pagesorchestrator.sandbox.uffd.startup.source_pagesorchestrator.sandbox.uffd.startup.bytesAttributes:
start_type∈{create, resume},success∈{true, false}. Nohigh-cardinality labels. Counts demand faults only (resolved, no
double-counting); prefaulted pages bypass the serve loop