Skip to content

Releases

SlowShield 0.0.13

Choose a tag to compare

@github-actions github-actions released this 25 Oct 09:05
fdaf55d

Added

  • Composer (PHP): Packagist through SlowShield at /composer/ (docs/design/composer.md).
    A version's publish time is Packagist's published-time, never the author's time; one in the future counts as
    now. The time belongs to the version's commit: a branch's new commit, or a tag moved on a mirror, is timed afresh.
    Versions younger than the delay are left out of the metadata, and a lock file that pins one gets 425 Too Early
    with a warning Composer prints. Every dist comes through SlowShield over HTTPS, checked against its first
    download; GitHub archives built again with the same files are not tampering, and an archive that can't be read is
    refused. A dist is judged by every version on its commit. Malware from OSV and GitHub is refused with 451,
    Packagist's malware list and security advisories reach Composer, and composer audit works. Brand-new packages
    are held too (upstreams.composer.fail_open = false). Needs Composer 2.10 or later. The Setup page has the
    commands, a composer.json, and a line that points a lock written against packagist.org at SlowShield;
    slowshield.org has a PHP guide.
  • RubyGems (Ruby) at /rubygems/ (#17): Bundler uses SlowShield as a
    mirror of rubygems.org (bundle config set --global mirror.https://rubygems.org <public_url>/rubygems), gem as
    its only source in ~/.gemrc; Gemfiles and Gemfile.lock keep naming rubygems.org. Each version and platform is
    judged by its own created_at, so a platform gem pushed late for an old version waits too. Held and blocked
    versions are left out of the compact index; a download of one gets 425 Too Early or 451. Every .gem is
    checked against the index's sha256 and its first-seen fingerprint, and every .gemspec.rz against its first-seen
    fingerprint. Brand-new gems are held (upstreams.rubygems.fail_open = false). gem search, gem outdated and
    gem update need rubygems.org's full index, which SlowShield doesn't serve. The Setup page has Bundler's mirror
    and its cooldown (Bundler 4.0.13 and later), and the .gemrc; Helm: ecosystems.rubygems.enabled; the website
    has a Ruby guide.

Security

  • Package-name checks no longer accept a name with a trailing newline, in every ecosystem. Before, rails\n passed
    as a gem name, and the blocklist, which names rails, didn't match it.
  • OpenTelemetry 1.45.1 (GHSA-xcq7-p3pf-p88h: unbounded tracestate parsing) and certifi 2026.10.19
    (GHSA-6q48-6c7w-9gw2: emSign roots with compromised intermediates). certifi is the CA bundle the OTLP exporter
    uses for HTTPS; package downloads use the system store, which distrusts emSign since 0.0.10. The website's
    Wrangler gets undici 7.30.1 (GHSA-xjj6-f33q-929c).

Images

  • ghcr.io/squirro/slowshield:0.0.13
  • ghcr.io/squirro/slowshield-caddy:0.0.13

Agent Skill

  • slowshield-skill.zip, sha256 4b3c9f0d6bad64a659b68fbb91480f20c9f17651c51dc619b6e7e4675249eb0d

SlowShield 0.0.12

Choose a tag to compare

@github-actions github-actions released this 25 Oct 00:09
86d073d

Added

  • Shield wall: several SlowShield instances working as one, a leader and its followers
    (docs/design/shieldwall.md, #34). Run in
    production across three regions, on hosts and on Kubernetes with the Helm chart.
    • Pairing: slowshield wall invite on the leader prints a join string (single use, 10 minutes, --minutes up to
      60); a follower started with it (SLOWSHIELD_JOIN) checks the leader's key and asks for a click on Join on its
      Shield wall page, or joins by itself with SLOWSHIELD_JOIN_CONFIRM=auto. Every message is signed with the
      instances' Ed25519 keys. A follower that has joined ignores a join string, so automation can leave it set
      (docs/operations.md).
    • An instance keeps its role across restarts: a follower syncs with the leader it stored, and only
      SLOWSHIELD_SHIELDWALL_ROLE=standalone takes an instance out of the shield wall.
    • The leader's overview, security timeline and CSV cover every instance, and narrow to one instance (the leader
      included), a location or a label. Followers report their statistics and events exactly once, including their
      history at pairing and what they queued while the leader was away.
    • The leader's Shield wall page lists every instance with its location, labels, version and last contact, and what
      each served in the last 24 hours: installs, packages, data, and how much came from its own cache, through the
      leader, or from the registries.
    • Followers take the leader's policy and blocks: stricter settings at once, looser ones after an hour and never
      below SLOWSHIELD_SHIELDWALL_MIN_DELAY_DAYS (1 day); a block the leader lifts stays a day. Their own stricter
      settings and their own exceptions stay in force. Takedowns, tamper flags, tag history and Maven, Cargo and NuGet
      listing times come down too, with the leader's observation times bounded by how recently the follower synced.
    • Package files come through the leader's cache when it is up (SLOWSHIELD_SHIELDWALL_VIA_LEADER), checked by the
      follower as always, so the wall downloads each file from the registries once; a follower that can't reach its
      leader works on its own.
    • The leader compares every follower's first fingerprint of a file with its own; a follower that saw different
      bytes refuses the file, and the leader records it. A follower's fingerprint only ever refuses a file on that
      follower, and one that names another package or version than the leader's record of the file is refused.
    • A follower must trust its leader's certificate, and the shipped stacks default to Caddy's private CA: Helm
      shieldwall.leaderCaSecret (the chart refuses a follower without it, unless shieldwall.leaderPubliclyTrusted),
      Compose LEADER_CA_SECRET_FILE. A follower that lacks it says so on its Shield wall page.
    • The leader sends fail_open for every ecosystem it has. A follower skips the settings for an ecosystem it
      doesn't know, with a warning that names it, and applies the rest of the policy.
    • Settings in a follower's own config count against its leader's policy (stricter wins), so the shipped configs
      leave default_delay_days, enforce_age_on_download and fail_open commented out.
  • One Tailscale Service name for many instances, on hosts (tailscale serve --service) and on Kubernetes, and a Helm
    follower of a leader that is only on a tailnet: docs/tailscale.md, with example manifests in
    deploy/helm/slowshield/examples.

Images

  • ghcr.io/squirro/slowshield:0.0.12
  • ghcr.io/squirro/slowshield-caddy:0.0.12

Agent Skill

  • slowshield-skill.zip, sha256 1178122df1f248a2a0ab0310dc762c05363fab0d1b086f8f9c4c1755f073db9d

SlowShield 0.0.11

Choose a tag to compare

@github-actions github-actions released this 24 Oct 16:42
d4f72e9

Added

  • setup_url (SLOWSHIELD_SETUP_URL): the URL the Setup page gives clients, when it isn't public_url. Moving
    clients to a new name (a Tailscale Service, say), an instance keeps answering under public_url and keeps its
    npm, Cargo and NuGet URLs there, so existing clients go on working, while new ones are set up with the new name.
  • Helm: hostAliases for the pod, for a host the cluster can't resolve, such as one only on a tailnet: its name
    can point at an in-cluster egress proxy (a Tailscale egress Service, say) and keep its URL and certificate.

Fixed

  • UI: the filter dropdowns match the design, in light and dark mode; before, they had the browser's native look.

Images

  • ghcr.io/squirro/slowshield:0.0.11
  • ghcr.io/squirro/slowshield-caddy:0.0.11

Agent Skill

  • slowshield-skill.zip, sha256 c4bd87cf3ceb0937a1fddf7c14c42ee30d8ac4ea584b41617828d3758a11a295

SlowShield 0.0.10

Choose a tag to compare

@github-actions github-actions released this 24 Oct 12:42
7515aa7

Changed

  • Python 3.15.0, the stable release, instead of 3.15.0rc2, in the images (GIL and free-threaded), CI and
    .python-version. The patch version is pinned, so a newer uv no longer changes the interpreter.

Security

  • The images no longer trust emSign ECC Root CA - G3, emSign ECC Root CA - C3 and emSign Root CA - C1, which Amazon
    Linux 2023 still ships. Eight intermediates under G3 have compromised keys (GHSA-6q48-6c7w-9gw2), and Mozilla has
    dropped all three. Before, an on-path attacker holding one of those keys could have intercepted SlowShield's
    upstream requests and Caddy's ACME requests. The image build fails if one of them is still in the bundle.
  • uv 0.13.4 in the image builder and CI. Earlier versions trust emSign ECC Root CA - G3 in their bundled root
    certificates, and eight intermediates under it have compromised keys (GHSA-hq85-5cpj-mr62, critical). Taken
    inside the 7-day dependency cooldown as a security exception.
  • SlowShield no longer asks registries for Brotli-compressed metadata, so it never decodes a Brotli stream. It
    was not affected by CVE-2026-111497 ("Node2Shell", in the C Brotli decoder): its decoder is the Rust port in
    pyreqwest, and the images ship no libbrotli. Registries send gzip instead. A metadata or npm audit response in
    an encoding SlowShield didn't ask for is refused as an upstream error instead of being parsed or passed on.

Images

  • ghcr.io/squirro/slowshield:0.0.10
  • ghcr.io/squirro/slowshield-caddy:0.0.10

Agent Skill

  • slowshield-skill.zip, sha256 29233cf4bb71cd5270a9129bbbb99b67313d299caae58d7f3ea39adfc97ff438

SlowShield 0.0.9

Choose a tag to compare

@github-actions github-actions released this 10 Oct 18:51
10cc8eb

Added

  • NuGet (C#) at /nuget/: nuget.org through SlowShield (docs/design/nuget.md,
    #38). Clients use a NuGet.Config with <clear/> and one source
    named nuget.org at /nuget/v3/index.json. The Setup page has the file, a command for CI and images, and a
    Directory.Build.props that fails the build on NU1603 (a held version skipped for a higher one). The guide has
    a page for C#: https://slowshield.org/docs/csharp/.
    • SlowShield writes the service index itself: the flat container, one SemVer 2 registration hive, vulnerability
      data (NuGetAudit keeps working) and search. The upstream URLs come from [upstreams.nuget], so the hosts it may
      reach stay fixed.
    • Versions younger than the delay or blocked are left out of the flat container, the registration pages
      (recounted and rebounded, empty pages dropped) and search results. A download of one gets 425 Too Early
      with Retry-After, or 451. A version's publish time is its registration published, never earlier than
      first seen; an unlisted version that never had a real date is timed from when SlowShield first listed it.
    • A registration is kept as nuget.org sent it and a page decoded only to serve it, so the largest (Uno.WinUI,
      5,451 versions, 115 MB) take about their size in memory; one over 256 MB is refused with a 503.
    • Every .nupkg is checked against the catalog's packageHash (SHA512) and its first fingerprint, and never
      changed, so the repository signature stays valid. A package nuget.org signs again is refused as tampering, cached
      or not.
    • Ids and versions use nuget.org's spellings (lower case, normalized: 1.0 and 1.0.0.0 are 1.0.0) in every
      key, exceptions and blocks included; other spellings in a path are a 404. Malware from OSV NuGet and
      GitHub nuget. fail_open is off, as for Maven and Cargo. SLOWSHIELD_NUGET_ENABLED, Helm
      ecosystems.nuget.enabled.

Changed

  • The shipped configs (Compose, Podman, Helm) leave default_delay_days, enforce_age_on_download and fail_open at
    their defaults, commented out.
  • A sha512 mismatch is reported as "sha512 does not match the registry digest" (npm and NuGet), not as npm's
    dist.integrity.

Security

  • A file whose registry digest changed since it was first served (npm's dist.integrity, the sha256 on PyPI and
    crates.io) is refused as tampering even when a verified copy is cached; before, the cached copy was served.

Fixed

  • Caddy capped npm audit requests at 16 KB instead of 10 MB: every matching request_body limit applied, the
    catch-all's too.

Images

  • ghcr.io/squirro/slowshield:0.0.9
  • ghcr.io/squirro/slowshield-caddy:0.0.9

Agent Skill

  • slowshield-skill.zip, sha256 04d4151a51b5cc551f657b382470b5b8ab8cc87b744ca9ff398a6434dc506131

SlowShield 0.0.8

Choose a tag to compare

@github-actions github-actions released this 07 Oct 20:12
4cccb9a

Added

  • Container images at /v2/, the OCI distribution API, for pulls (docs/design/oci.md,
    #22). Built in: docker.io, ghcr.io, quay.io, registry.k8s.io, gcr.io,
    mcr.microsoft.com and public.ecr.aws.
    • Tags lag behind: a tag resolves to the newest digest it has pointed to for the delay, so nginx:latest keeps
      working, about a week behind. A tag's time is when it got that digest: from the Docker Hub API, Quay's tag
      history, Artifact Registry's upload times or MCR's catalog where the registry has one, and SlowShield's own
      first sight, whichever is earlier.
    • A pinned digest that is too new, a blocked image and a digest the registry took down are refused with 403 and
      a message Docker, Podman, BuildKit, skopeo and crane print (containerd 2.2 and older show only the status).
    • During an instance's first default_delay_days, tags it has no history for are served at their current digest
      and recorded as fail-open; upstreams.oci.fail_open = false is strict from the start.
    • Manifests are checked against their digests and kept by digest; blobs are checked while streaming. Layers aren't
      stored unless upstreams.oci.layer_cache_gb is above 0, which keeps them in a separate store (Helm:
      persistence.ociLayers). Redirects go only to each registry's download_hosts, without the registry token.
    • Docker Hub pulls: tags are resolved with HEAD, which doesn't count, and each manifest is fetched once for all
      clients. username and token_file add a Docker Hub token; slowshield_oci_ratelimit_remaining shows what is
      left.
    • The Setup page has containerd and Kubernetes, Docker, Podman (also CRI-O, Buildah and skopeo), BuildKit and
      image names, checked with the real clients: containerd, Docker's containerd image store and Podman don't go
      around SlowShield; BuildKit and Docker's classic store pull from the registry after a refusal, which the page
      says. The package page shows every digest each tag pointed to.
  • Operator blocks for every ecosystem: [[blocks]] in config.toml refuses a package, a version, or an image
    repository, tag or digest, like a malware advisory. They apply on reload and are listed with source config.
  • The Setup page links every tool to its guide on slowshield.org.

Changed

  • The UI names ecosystems by language, as slowshield.org does: Python, JavaScript, Go, Java, Rust and Containers
    instead of PyPI, npm, Go, Maven, Cargo and OCI. URLs, config, metrics and CSV exports keep the ids (pypi, npm,
    …).
  • Digests are shortened wherever the UI shows a version (sha256:3734a9c4892e…), with the full value on hover.

Fixed

  • Links from the security events to a package that was only ever refused, never served, led to "not found"; such a
    package now has a page.

Images

  • ghcr.io/squirro/slowshield:0.0.8
  • ghcr.io/squirro/slowshield-caddy:0.0.8

SlowShield 0.0.7

Choose a tag to compare

@github-actions github-actions released this 06 Oct 19:16
2a8339b

Added

  • Cargo at /cargo/: a sparse registry that replaces crates-io in $CARGO_HOME/config.toml
    (docs/design/cargo.md, #16). Cargo.lock keeps
    crates.io as its source and checksums. Needs cargo 1.68 or later.

    • Index files pass through byte for byte, except that versions too new or blocked are marked
      "yanked":true, so cargo resolves to an older version. A Cargo.lock that pins one gets 403 with the time it
      becomes available and the cargo update --precise command for the newest version allowed; malware gets 451.
      Cargo prints both. Upstream failures are 503, which cargo retries.
    • A version's publish time is its pubtime in the index, which crates.io sets. The first one seen is kept, so a
      rewritten index can't move it earlier. No requests to the crates.io API.
    • Downloads are checked against the index's cksum (sha256) and fingerprinted; static.crates.io is always asked
      for the index's exact spelling of the name.
    • Crate names match in crates.io's canonical form (lower case, - as _) in exceptions, advisories and the UI.
    • fail_open is off for Cargo: crates none of whose versions is old enough are held too.
  • The OSV feed also takes RustSec's malware advisories (category malicious) that have no MAL-* counterpart: most
    crates RustSec reports as malware never got one. GitHub's malware advisories for Rust are read too.

  • The Setup page has Cargo: the config.toml, and a one-line command for CI and Dockerfiles (the official rust
    images set CARGO_HOME, so ~/.cargo isn't read there).

  • The Setup page also sets the package managers' own release age, 3 days, where they have one (pip, uv, Poetry,
    PDM, npm, pnpm, Yarn, Bun), as a second layer: a release then still waits on a machine that goes around
    SlowShield, or when SlowShield serves a brand-new package because nothing is old enough yet (fail-open). In
    normal use only SlowShield holds anything, since its delay is longer. A switch turns it off; each tool shows the
    version it needs, and what older versions do with the setting. Never more than SlowShield's own delay. uv's and
    PDM's go in pyproject.toml only: uv records the setting in uv.lock (an environment variable on one machine
    breaks uv sync --locked on another), and PDM doesn't keep its command-line flag.

Changed

  • Cargo is shown in #DEA584 (GitHub's Rust colour) in the UI and in Grafana.

Images

  • ghcr.io/squirro/slowshield:0.0.7
  • ghcr.io/squirro/slowshield-caddy:0.0.7

SlowShield 0.0.6

Choose a tag to compare

@github-actions github-actions released this 06 Oct 11:06
f5c00fe

Added

  • Maven repositories at /maven/ for Maven, Gradle, sbt and Coursier (docs/design/maven.md,
    #18). /maven/all/ serves Maven Central and Google Maven behind one
    URL for a settings.xml mirror; /maven/central/, /maven/google/ and /maven/gradle-plugins/ serve each
    repository (Gradle's init script uses those), and operator repositories go under /maven/<id>/.
    • maven-metadata.xml leaves out versions that are too new or blocked, with <latest>/<release> and the
      checksum files recomputed. Files that are too new get 425 Too Early, which Maven and Gradle show as the reason
      and re-request on the next build. Malware gets 451; upstream failures 503, never a 404 Maven would cache.
    • A file's publish time is its Last-Modified on the repository, or when SlowShield first saw the version listed,
      whichever is earlier, so a file added to an old version later is held on its own.
    • Downloads are checked against Central's x-checksum-sha1, the .sha1 file, or the sha256 in the Plugin
      Portal's path, and fingerprinted. Responses carry X-Checksum-Sha1, so Maven skips checksum requests.
    • OSV and GitHub malware advisories for Maven feed the blocklist, matched in Maven's version order (1.0 = 1.0.0).
    • Metadata SlowShield can't filter is refused with 503, never passed on unfiltered.
    • A cached file is judged again by its own recorded Last-Modified (migration 0004_artifact_published), so it is
      held again when the policy gets stricter, without a request upstream.
  • fail_open per ecosystem (upstreams.<ecosystem>.fail_open). Maven defaults to off: brand-new artifacts are held.
  • The Setup page has Maven (settings.xml), Gradle (init script), sbt and Coursier.

Changed

  • Maven is shown in Java orange #ED8B00 (Maven's own red is too close to npm's), in the UI and in Grafana.
  • CI: the performance comparison on pull requests (k6 against the base branch, and the micro benchmarks) can be
    skipped with the skip-perf label, and release pull requests (branches release-*) skip it. Releases no longer
    run a performance gate against the previous release, only the end-to-end tests and the observability smoke test
    on the release images.

Images

  • ghcr.io/squirro/slowshield:0.0.6
  • ghcr.io/squirro/slowshield-caddy:0.0.6

SlowShield 0.0.5

Choose a tag to compare

@github-actions github-actions released this 05 Oct 15:45
ca0cda4

Added

  • Go modules at /go/: set GOPROXY=https://<host>/go (without ,direct). Versions younger than the delay are left
    out of version lists, and requests for them get 403 with a Retry-After header and a message the go command
    prints. Known malware gets 451, never 404/410, so a ,direct fallback can't go around SlowShield either.
    .mod and .zip downloads are checked against the h1: hashes in sum.golang.org and against their first-seen
    fingerprint. The checksum database itself is proxied at /go/sumdb/sum.golang.org/, so clients need no other
    route out. A version's publish time is when proxy.golang.org first stored it (the Last-Modified of its .mod),
    never the commit time, which authors can backdate. It is looked up once per version, at no extra load on the
    upstreams beyond one HEAD (docs/design/go.md,
    #15). @latest of a module without tags answers with the newest
    commit known to be old enough when the newest one is too new. OSV and GitHub malware advisories for Go feed the
    blocklist.
    Configuration: [upstreams.go], SLOWSHIELD_GO_ENABLED, Helm ecosystems.go.enabled.
  • The Setup page and slowshield.org set GOPROXY in the shell setup, and the tool finder has Go.
  • A nightly check that proxy.golang.org's Last-Modified still matches index.golang.org, and Go perf scenarios
    (go_list, go_mod). The perf gate reports a scenario the baseline release doesn't serve yet without comparing it.

Changed

  • Ecosystems are shown in their logos' primary colours, in the UI and the Grafana dashboards: PyPI #3775A9, npm
    #CB3837, Go #00ADD8, the same in light and dark mode.

Fixed

  • Chart axes in the UI show round, distinct values, with each gridline at its value. Small counts used to repeat
    labels (a peak of 2 read 0, 1, 2, 2).

Images

  • ghcr.io/squirro/slowshield:0.0.5
  • ghcr.io/squirro/slowshield-caddy:0.0.5
Performance (amd64)

SlowShield performance report

Host: Linux/x86_64, 4 CPUs · candidate ghcr.io/squirro/slowshield@sha256:73cc215262b653e60c517b5346671a7775304f276699e4ca94784549bf3c459e · baseline ghcr.io/squirro/slowshield:latest

Scenario Metric Baseline Candidate Change Limit p
artifact_big:throughput rps 670 637 -5.0% 15% 0.690 ✅
artifact_cached:latency p95_ms 0.73 0.73 -0.0% 15% 0.841 ✅
artifact_cached:latency p99_ms 0.93 0.94 +1.1% 20% 0.548 ✅
artifact_cached:throughput rps 2583 2581 -0.1% 10% 0.690 ✅
blocked:latency p95_ms 0.57 0.57 -0.0% 15% 0.690 ✅
blocked:latency p99_ms 0.79 0.76 -3.5% 20% 0.421 ✅
dashboard:latency p95_ms 2.37 2.45 +3.2% 15% 0.016 ✅
dashboard:latency p99_ms 2.85 3.00 +5.3% 30% 0.548 ✅
mixed:throughput rps 3875 3873 -0.0% 10% 0.690 ✅
npm_huge_corgi:throughput rps 558 562 +0.7% 10% 1.000 ✅
npm_huge_full:throughput rps 431 431 +0.1% 10% 0.841 ✅
npm_packument_corgi:throughput rps 5174 5178 +0.1% 10% 0.548 ✅
npm_packument_full:latency p95_ms 0.46 0.46 -0.1% 15% 0.421 ✅
npm_packument_full:latency p99_ms 0.69 0.70 +0.7% 20% 1.000 ✅
pypi_simple_html:latency p95_ms 1.37 1.30 -5.2% 15% 1.000 ✅
pypi_simple_html:latency p99_ms 1.94 1.98 +1.9% 20% 1.000 ✅
pypi_simple_json:latency p95_ms 1.86 2.04 +9.2% 15% 1.000 ✅
pypi_simple_json:latency p99_ms 3.03 3.25 +7.2% 20% 0.690 ✅
pypi_simple_json:throughput rps 1029 1055 +2.5% 10% 0.548 ✅
process rss_peak_bytes 173.3 MB 173.6 MB +0.1% 15% 0.690 ✅
process startup_s 0.55 0.55 +0.3% 30% 0.841 ✅
image size_bytes 124.1 MB 124.2 MB +0.1% 5% — ✅

Candidate medians

Scenario rps p50 ms p95 ms p99 ms MB/s CPU ms/1k
artifact_big:throughput 637 0.71 0.71 0.71 2.1 1575.0
artifact_cached:latency 300 0.63 0.73 0.94 0.5 495.9
artifact_cached:throughput 2581 11.98 17.79 23.25 4.0 387.2
blocked:latency 300 0.48 0.57 0.76 0.2 333.0
dashboard:latency 30 1.89 2.45 3.00 0.5 1901.6
go_list:throughput 4177 7.20 12.23 17.97 18.4 237.9
go_mod:latency 300 0.73 0.88 1.32 0.2 622.8
mixed:throughput 3873 6.68 16.65 21.19 9.2 256.9
npm_huge_corgi:throughput 562 13.64 21.49 25.54 1241.8 1778.0
npm_huge_full:throughput 431 18.23 27.10 31.20 1424.6 2315.9
npm_packument_corgi:throughput 5178 5.62 10.37 16.56 7.9 188.3
npm_packument_full:latency 300 0.37 0.46 0.70 0.7 249.8
pypi_simple_html:latency 300 0.87 1.30 1.98 110.4 686.9
pypi_simple_json:latency 300 1.11 2.04 3.25 138.1 953.7
pypi_simple_json:throughput 1055 28.48 44.70 54.08 485.9 949.6

Image size: 124.2 MB (uncompressed), RSS peak median 173.6 MB

Performance (arm64)

SlowShield performance report

Host: Linux/aarch64, 4 CPUs · candidate ghcr.io/squirro/slowshield@sha256:9e69760cca7008b2cfb5a174b73a1a1ac4eeee22b22a08d08da2c9ee76dc4789 · baseline ghcr.io/squirro/slowshield:latest

Scenario Metric Baseline Candidate Change Limit p
artifact_big:throughput rps 417 637 +52.8% 15% 0.095 ✅
artifact_cached:latency p95_ms 0.73 0.72 -1.5% 15% 0.222 ✅
artifact_cached:latency p99_ms 0.97 1.04 +6.9% 20% 0.548 ✅
artifact_cached:throughput rps 3082 3144 +2.0% 10% 0.151 ✅
blocked:latency p95_ms 0.55 0.54 -1.4% 15% 0.222 ✅
blocked:latency p99_ms 0.68 0.67 -2.4% 20% 0.690 ✅
dashboard:latency p95_ms 3.12 3.23 +3.3% 15% 0.095 ✅
dashboard:latency p99_ms 3.46 3.72 +7.5% 30% 0.056 ✅
mixed:throughput rps 4814 4911 +2.0% 10% 0.095 ✅
npm_huge_corgi:throughput rps 806 832 +3.3% 10% 0.548 ✅
npm_huge_full:throughput rps 569 586 +3.0% 10% 0.548 ✅
npm_packument_corgi:throughput rps 7428 7498 +0.9% 10% 0.056 ✅
npm_packument_full:latency p95_ms 0.42 0.43 +3.0% 15% 0.548 ✅
npm_packument_full:latency p99_ms 0.65 0.57 -12.3% 20% 0.095 ✅
pypi_simple_html:latency p95_ms 1.34 1.32 -1.6% 15% 1.000 ✅
pypi_simple_html:latency p99_ms 1.73 1.72 -0.3% 20% 0.548 ✅
pypi_simple_json:latency p95_ms 1.71 1.57 -7.8% 15% 0.841 ✅
pypi_simple_json:latency p99_ms 2.59 2.12 -18.2% 20% 0.690 ✅
pypi_simple_json:throughput rps 1571 1581 +0.6% 10% 0.690 ✅
process rss_peak_bytes 153.4 MB 153.9 MB +0.3% 15% 1.000 ✅
process startup_s 0.54 0.53 -2.3% 30% 0.151 ✅
image size_bytes 115.9 MB 116.0 MB +0.1% 5% — ✅

Candidate medians

Scenario rps p50 ms p95 ms p99 ms MB/s CPU ms/1k
artifact_big:throughput 637 0.82 0.82 0.82 2.1 1204.0
artifact_cached:latency 300 0.66 0.72 1.04 0.5 578.1
artifact_cached:throughput 3144 9.99 14.60 18.14 4.9 317.8
blocked:latency 300 0.50 0.54 0.67 0.2 346.5
dashboard:latency 30 2.43 3.23 3.72 0.5 2325.4
go_list:throughput 5672 5.40 9.02 11.65 25.1 175.8
go_mod:latency 300 0.78 0.87 1.06 0.2 705.6
mixed:throughput 4911 5.60 13.17 16.02 11.7 203.5
npm_huge_corgi:throughput 832 9.29 13.93 16.50 1838.9 1201.1
npm_huge_full:throughput 586 13.51 18.70 21.80 1937.2 1704.6
npm_packument_corgi:throughput 7498 3.83 7.36 11.29 11.5 131.4
npm_packument_full:latency 300 0.36 0.43 0.57 0.7 270.1
pypi_simple_html:latency 300 0.90 1.32 1.72 110.4 786.1
pypi_simple_json:latency 300 0.91 1.57 2.12 138.2 844.0
pypi_simple_json:throughput 1581 19.32 29.28 35.80 728.0 632.0

Image size: 116.0 MB (uncompressed), RSS peak median 153.9 MB

SlowShield 0.0.4

Choose a tag to compare

@github-actions github-actions released this 04 Oct 18:22
b480db5

Added

  • The Setup page follows slowshield.org: one global shell setup (bash on Linux, bash on macOS, zsh, fish) filled in
    with this instance's URLs and preselected from the visitor's OS, a one-line install to try it, a tool finder (type
    poe for Poetry, ya for Yarn) for pip, uv, Poetry, PDM, Pipenv, npm, pnpm, Yarn and Bun, and CI/Dockerfile
    snippets. Its shell and install snippets come from the same file as the website's
    (src/slowshield/ui/snippets.toml), and an end-to-end test runs every one of them in bash, zsh and fish against
    the stack, including Debian 12's pip 23.0.1.

Changed

  • One host, one path per ecosystem (docs/design/routing.md). The UI lives entirely
    under /ui/: the dashboard moved from / to /ui/ (/ redirects there) and assets from /static/ to
    /ui/static/. The root paths SlowShield answers on are a fixed list, with names reserved for every ecosystem
    on the roadmap.
  • The Setup page shows path URLs only (/pypi/simple/, /npm/).

Deprecated

  • Per-ecosystem hostnames (upstreams.pypi.hostnames / SLOWSHIELD_PYPI_HOSTNAMES, upstreams.npm.hostnames /
    SLOWSHIELD_NPM_HOSTNAMES, Helm ecosystems.*.hostnames), the root PyPI alias (/simple/, /packages/) and
    /static/. They keep working until 0.1, log a startup warning (hostnames), show a notice on the Setup page,
    and are counted in slowshield_legacy_routing_requests_total so operators can see when nothing uses them.

Security

  • Values shown in copy-paste shell snippets can no longer carry shell syntax (reported by Aikido). With
    local_http on, a loopback Host header such as localhost:$(id) was echoed into the Setup page and npm
    tarball links; only well-formed hosts with a numeric port qualify now. public_url and
    upstreams.npm.public_url must be plain http(s)://host[:port][/path] URLs (no query, credentials or
    special characters), and the snippet renderer refuses anything else. A config with such a URL now fails to
    load instead of being served.

Fixed

  • The UI's CSS and JavaScript URLs are versioned by content instead of the release, so an image rebuilt under the
    same version no longer leaves browsers on stale assets.
  • With npm hostnames configured, clients using /npm/ got tarball links on the npm hostname, which then ended
    up in their lockfiles. Tarball links now follow the route the client used.

Images

  • ghcr.io/squirro/slowshield:0.0.4
  • ghcr.io/squirro/slowshield-caddy:0.0.4
Performance (amd64)

SlowShield performance report

Host: Linux/x86_64, 4 CPUs · candidate ghcr.io/squirro/slowshield@sha256:aa16651770ecfb35cc2143f1ed9926c3d35c17b65d8722573d1c50ce2d510cb8 · baseline ghcr.io/squirro/slowshield:latest

Scenario Metric Baseline Candidate Change Limit p
artifact_big:throughput rps 691 678 -2.0% 15% 0.690 ✅
artifact_cached:latency p95_ms 0.70 0.71 +1.3% 15% 0.016 ✅
artifact_cached:latency p99_ms 0.84 0.86 +3.3% 20% 0.421 ✅
artifact_cached:throughput rps 2547 2540 -0.3% 10% 0.690 ✅
blocked:latency p95_ms 0.57 0.58 +0.1% 15% 1.000 ✅
blocked:latency p99_ms 0.77 0.75 -3.3% 20% 0.690 ✅
dashboard:latency p95_ms 2.36 2.38 +1.0% 15% 0.151 ✅
dashboard:latency p99_ms 3.02 2.98 -1.3% 30% 0.421 ✅
mixed:throughput rps 3845 3827 -0.5% 10% 0.841 ✅
npm_huge_corgi:throughput rps 551 560 +1.6% 10% 0.151 ✅
npm_huge_full:throughput rps 432 441 +2.0% 10% 0.421 ✅
npm_packument_corgi:throughput rps 5164 5136 -0.5% 10% 0.310 ✅
npm_packument_full:latency p95_ms 0.45 0.45 -0.4% 15% 1.000 ✅
npm_packument_full:latency p99_ms 0.61 0.63 +2.5% 20% 0.690 ✅
pypi_simple_html:latency p95_ms 1.32 1.80 +35.7% 15% 0.151 ✅
pypi_simple_html:latency p99_ms 1.95 2.29 +17.4% 20% 0.151 ✅
pypi_simple_json:latency p95_ms 1.90 2.07 +8.9% 15% 0.548 ✅
pypi_simple_json:latency p99_ms 2.68 4.05 +51.0% 20% 0.222 ✅
pypi_simple_json:throughput rps 1023 964 -5.8% 10% 0.690 ✅
process rss_peak_bytes 168.1 MB 171.1 MB +1.8% 15% 0.151 ✅
process startup_s 0.56 0.56 +0.5% 30% 0.222 ✅
image size_bytes 124.0 MB 124.1 MB +0.0% 5% — ✅

Candidate medians

Scenario rps p50 ms p95 ms p99 ms MB/s CPU ms/1k
artifact_big:throughput 678 0.80 0.80 0.80 2.2 1684.0
artifact_cached:latency 300 0.65 0.71 0.86 0.5 499.1
artifact_cached:throughput 2540 12.22 17.99 23.76 4.0 393.5
blocked:latency 300 0.49 0.58 0.75 0.2 337.7
dashboard:latency 30 1.76 2.38 2.98 0.4 1782.7
mixed:throughput 3827 6.74 16.85 21.24 9.1 259.4
npm_huge_corgi:throughput 560 13.65 21.62 25.58 1237.9 1777.0
npm_huge_full:throughput 441 17.92 26.54 30.48 1456.3 2265.9
npm_packument_corgi:throughput 5136 5.68 10.57 16.63 7.9 189.9
npm_packument_full:latency 300 0.37 0.45 0.63 0.7 248.6
pypi_simple_html:latency 300 1.09 1.80 2.29 110.4 867.2
pypi_simple_json:latency 300 1.13 2.07 4.05 138.1 975.7
pypi_simple_json:throughput 964 31.25 49.45 59.47 443.9 1036.2

Image size: 124.1 MB (uncompressed), RSS peak median 171.1 MB

Performance (arm64)

SlowShield performance report

Host: Linux/aarch64, 4 CPUs · candidate ghcr.io/squirro/slowshield@sha256:1633ac14fe1e7b7e350bf1ccdaae48e5beb4f56a83a19f4c32b672beaaace186 · baseline ghcr.io/squirro/slowshield:latest

Scenario Metric Baseline Candidate Change Limit p
artifact_big:throughput rps 403 461 +14.4% 15% 0.310 ✅
artifact_cached:latency p95_ms 0.73 0.70 -4.7% 15% 1.000 ✅
artifact_cached:latency p99_ms 0.84 0.84 -0.8% 20% 1.000 ✅
artifact_cached:throughput rps 3071 3158 +2.8% 10% 0.421 ✅
blocked:latency p95_ms 0.60 0.57 -4.4% 15% 0.310 ✅
blocked:latency p99_ms 0.77 0.76 -1.8% 20% 0.548 ✅
dashboard:latency p95_ms 3.24 3.08 -4.8% 15% 0.222 ✅
dashboard:latency p99_ms 3.67 3.43 -6.4% 30% 0.310 ✅
mixed:throughput rps 4831 4921 +1.9% 10% 0.222 ✅
npm_huge_corgi:throughput rps 696 826 +18.7% 10% 0.548 ✅
npm_huge_full:throughput rps 560 575 +2.5% 10% 0.690 ✅
npm_packument_corgi:throughput rps 7348 7520 +2.3% 10% 0.222 ✅
npm_packument_full:latency p95_ms 0.42 0.41 -3.6% 15% 0.548 ✅
npm_packument_full:latency p99_ms 0.59 0.56 -6.0% 20% 0.548 ✅
pypi_simple_html:latency p95_ms 1.31 1.22 -7.2% 15% 0.690 ✅
pypi_simple_html:latency p99_ms 1.88 1.77 -5.9% 20% 0.841 ✅
pypi_simple_json:latency p95_ms 1.67 1.68 +0.6% 15% 0.690 ✅
pypi_simple_json:latency p99_ms 2.44 2.04 -16.1% 20% 1.000 ✅
pypi_simple_json:throughput rps 1396 1545 +10.7% 10% 1.000 ✅
process rss_peak_bytes 152.7 MB 154.9 MB +1.5% 15% 1.000 ✅
process startup_s 0.53 0.52 -1.9% 30% 0.548 ✅
image size_bytes 115.8 MB 115.9 MB +0.0% 5% — ✅

Candidate medians

Scenario rps p50 ms p95 ms p99 ms MB/s CPU ms/1k
artifact_big:throughput 461 1.00 1.00 1.00 1.5 980.0
artifact_cached:latency 300 0.65 0.70 0.84 0.5 565.5
artifact_cached:throughput 3158 9.92 14.52 18.72 4.9 316.5
blocked:latency 300 0.51 0.57 0.76 0.2 360.0
dashboard:latency 30 2.30 3.08 3.43 0.4 2251.9
mixed:throughput 4921 5.62 13.22 15.89 11.7 203.0
npm_huge_corgi:throughput 826 9.33 14.08 16.89 1827.1 1209.1
npm_huge_full:throughput 575 13.87 19.09 21.70 1898.2 1739.5
npm_packument_corgi:throughput 7520 3.81 7.37 11.23 11.5 131.0
npm_packument_full:latency 300 0.35 0.41 0.56 0.7 269.1
pypi_simple_html:latency 300 0.97 1.22 1.77 110.4 835.6
pypi_simple_json:latency 300 1.01 1.68 2.04 138.2 902.3
pypi_simple_json:throughput 1545 19.75 30.40 37.09 711.4 646.7

Image size: 115.9 MB (uncompressed), RSS peak median 154.9 MB