The shipped examples had drifted from the code that parses them:
- configs/cluster.json was still in the pre-v3 format: no top-level "name",
"numberOfNodes" instead of the required "nodes" node-list expression, plain
numbers instead of MetricValue objects for flopRateScalar/flopRateSimd/
memoryBandwidth, unit strings instead of unit objects, "aggregation": null,
and hwthread 72 missing from the core topology. It neither validated against
cluster.schema.json nor decoded into schema.Cluster. Rewritten in the current
format and extended with the accelerator metrics the GPU subclusters need.
- archive.retention.location is not a field of taskmanager.Retention. That
decode is not strict, so the key was silently dropped and a "move" policy ran
with an empty target path. Replaced with target-kind/target-path.
- "ui-file" pointed at ui-config.json while the shipped file is uiConfig.json,
so ccConfig logged a load error and the UI fell back to defaults.
- Added the options introduced since the examples were written: checkpoint
file-format/interval/max-wal-size, api-subjects concurrency, db-config,
nodestate-retention, archive retention/compression, smoothing-window and the
resample policy/algo defaults.
The config written by -init had no metric-store section, so a server started
right after -init aborted with "missing metricstore configuration".
internal/configexample guards all of this: it resolves "-file" references the
way ccConfig does, strict-decodes every section into the struct that actually
parses it, and validates the cluster configs against the cc-lib schema.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
cc-lib's resamplers cannot smooth: validateFrequency bails out when the new
frequency is not coarser than the old one, and every algorithm produces an
output strictly shorter than its input. They consolidate for transport,
which is a separate job from filtering for display, and they only act at all
once a series exceeds the target point count - shorter jobs arrive raw.
Add a centered, NaN-aware moving average in the plot layer, applied after
whatever downsampling the backend performed. It preserves length and index
alignment, so the X array, the visible point count and the zoom-resampling
hook are unaffected. NaN and null samples are skipped, an all-NaN window
stays NaN so gaps keep rendering, and the window shrinks at the series edges
rather than introducing new gaps there.
The window is a user setting in data points, defaulting to 3, and is
orthogonal to the resample algorithm - it stacks on top of average
consolidation rather than replacing it. All series of a stats plot are
smoothed together, since smoothing only some would break the
min <= mid <= max invariant the plot bands rely on. Smoothing is display
only: reported statistics and job footprints come from JobMetric.Statistics
and are untouched. The uPlot cursor readout does show the smoothed value.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Average consolidation makes every plotted point the true arithmetic mean of
its interval, which is what the Ganglia-style plot appearance depends on.
LTTB deliberately keeps extremes instead, so a plot downsampled with it
cannot be read as interval means.
resolveResampleAlgo used to return an empty string when neither the user nor
the config had chosen an algorithm, and cc-lib's GetResampler maps that to
LTTB. Route every fallback through config.ResampleAlgo() and ship "average"
as the generated and example default.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
config.ResampleConfig stopped emitting the trigger and resolutions fields,
but MetricPlot still read them. resampleResolutions and resampleMinimum
were therefore always null, which made the array-based resolution branch in
the setScale hook unreachable and the timestep !== resampleMinimum guard
always true. Keep only the policy-based branch and derive the zoom trigger
straight from targetPoints.
configs/config-large.json still carried the removed minimum-points, trigger
and resolutions keys, which DisallowUnknownFields now rejects, so starting
against it aborted. Replace them with the current keys and add a test that
loads the file, so the example configs cannot drift out of the schema
unnoticed again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Only archived job data and internal-store node-list data honoured the
selected resample algorithm. For running jobs the algorithm was dropped:
MetricDataRepository.LoadData had no such parameter, so the memory store
fell back to an empty string, which cc-lib's GetResampler maps to LTTB.
The external store client was worse - its APIQueryRequest had no
ResampleAlgo field at all, and LoadNodeListData accepted the parameter
without using it.
Add resampleAlgo to the LoadData interface (mirroring LoadNodeListData),
forward it from metricdispatch, and set it on both stores' requests. The
field is tagged omitempty, so the wire format is unchanged when empty -
verify the deployed cc-metric-store accepts it before relying on it there.
The REST job endpoints pass a non-zero resolution and therefore do
resample, so they now request the configured default instead of an empty
string. Add config.ResampleAlgo() for that, since "" is not a neutral
value at this layer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The policy -> target-points mapping existed twice with different values:
internal/config (300/600/1000) fed the resampler's MinimumRequiredPoints
threshold, while internal/metricdispatch (200/500/1000) fed the requested
resolution and the frontend target point count.
Because the threshold was larger than the target, the resampler refused to
downsample series whose length fell between the two numbers, silently
dropping the resolution the backend had asked for. With the medium policy
that covered every series between 500 and 600 points.
Move the table into internal/config as the single source of truth (import
direction rules out the reverse, since metricdispatch already imports
config) and keep the 200/500/1000 values, which already drove the requested
resolution. metricdispatch.TargetPointsForPolicy now delegates to it, so
MinimumRequiredPoints equals the target and resampling happens exactly when
a series exceeds it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Pre-existing formatting drift picked up by `make fmt`. Whitespace and
comment layout only, no semantic change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Resolve conflicts in the generated GraphQL code by regenerating it against
the merged schema. cc-lib v2.13.0 adds Tooltip to schema.MetricConfig and
schema.GlobalMetricListItem, so gqlgen now binds the tooltip field directly
and the hand-written globalMetricListItem/metricConfig resolvers introduced
on the tooltip branch are no longer needed.
Also migrate to the cc-lib v2.13.0 metric container types, which changed
from bare maps to structs carrying array-valued metric groups:
schema.JobData map -> {Metrics, Groups}
schema.ScopedJobStats map -> {Metrics, Groups}
job.Statistics map -> schema.JobStatisticsSet{Metrics, Groups}
Callers index .Metrics, return the zero struct instead of nil, and
deepCopy/DecodeJobStats now also carry the Groups payload through.
archive.GetStatistics returns the full JobStatisticsSet so group
statistics survive the round trip.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
PlotRenderOptions was rendered both in AdminSettings.svelte and in
UserSettings.svelte. Since Config.root.svelte always renders UserSettings,
including for admins, admins saw the identical plot-render cards twice on
/config.
The admin copy was also misleading: every form in the component posts to
/frontend/configuration/, whose handler (RestAPI.updateConfiguration) always
writes per-user config via UserCfgRepo.UpdateConfig — never a global default.
On top of that, the admin-side handleSettingSubmit lacked the colorblind-mode
special case that UserSettings uses to keep PlotColorScheme in sync, so
toggling the option there did not refresh the palette preview.
Removes the component from AdminSettings.svelte along with the message,
displayMessage, handleSettingSubmit and popMessage members that existed only
to feed it. The remaining admin children manage their own messages.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reverts 22c442db "Enable entire integration" and drops the accompanying
.gitignore entry:
- .claude/settings.json: removes all seven "entire hooks claude-code ..."
command hooks (PostToolUse on Task and TodoWrite, PreToolUse on Task,
SessionEnd, SessionStart, Stop, UserPromptSubmit) and the
Read(./.entire/metadata/**) deny rule. The file held nothing else, so it
is deleted rather than left as an empty object.
- .entire/settings.json, .entire/.gitignore: untracked and removed.
- .gitignore: drops the now-pointless .entire entry.
.claude/settings.local.json is untouched; it carries unrelated local
permission settings.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
recomputeStats() wrote statSum/statSamples/statMin/statMax/statsValid from
inside the stats() fast path, which runs under only a shared RLock via
MemoryStore.Stats -> Level.findBuffers. Two concurrent queries on the same
invalid buffer could both enter the fast path and race on those fields.
Require statsValid in the fast-path guard instead of recomputing inline;
invalid buffers now fall through to the existing point-by-point scan, which
only reads b.data. recomputeStats() is unchanged and still used by
checkpoint loadFile at single-threaded load time.
Updated stats_test.go: overwritten buffers now stay statsValid=false after a
query (documenting non-mutating reads); TestStatsMultiBufferChain and
TestStatsFastPathThenPartialTail now recompute stats after building bare
buffers so the fast-path cache fold is still covered; added
TestStatsConcurrentQueriesNoRace (race-clean under -race for both valid and
post-overwrite buffers) and TestStatsGappedChain (real inter-buffer gap
correctly excluded from Samples/Min/Max/Avg).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add len(b.data) > 0 guard to prevent infinite loop when an empty linked
buffer is reachable. With the guard false, the normal path handles empty
buffers safely via the existing t < b.start || idx >= len(b.data) check.
Add regression test TestStatsFastPathThenPartialTail to pin the boundary:
three-buffer chain where earlier buffers are fully covered (fast path) and
the last buffer is partially covered (normal path).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
MemoryUsageTracker freed all buffers older than the retention cutoff on
every tick whenever the last retention pass had excluded used nodes,
regardless of memory pressure — discarding the history kept in memory
for long-running jobs. Buffers are now only reclaimed by the existing
emergency path once memory usage exceeds the configured cap. The
now-unused selectorsExcluded state is removed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Setting the provider after Init left the checkpoint restore inside Init
without job information and forced callers to pre-call InitMetrics and
SetNodeProvider in the right order. Init now receives the provider as a
parameter and injects it before the restore, so the ordering is enforced
by the signature. SetNodeProvider remains for callers that do not run
Init (tests, -cleanup-checkpoints).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The -cleanup-checkpoints CLI flag called CleanupCheckpoints without
initializing the MemoryStore singleton or setting its NodeProvider,
so it silently fell back to legacy clean-everything behavior and
could delete checkpoints for hosts with running jobs. Inject the job
repository as NodeProvider the same way runServer does, so the CLI
path also skips used hosts. Also update SetNodeProvider's doc comment,
which only mentioned Free but is now also consulted by FromCheckpoint
and CleanupCheckpoints.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
OIDC roles were translated by a hardcoded switch that only recognised the
literal names user/admin/manager/support (dropping api) and required the
IdP to emit exactly those names.
Add an optional auth.oidc.role-mapping (OIDC role/group claim value -> CC
role) so operators can map their own realm/client role names, including
api. The mapping is the sole source of roles: only mapped names are
honored, unmapped token roles are ignored, and users without any mapped
role receive the base "user" role. Mapping targets are validated once at
startup. Role assignment stays authoritative on every login as before.
Note: deployments relying on the IdP emitting literal CC role names must
now map them explicitly (e.g. "admin": "admin").
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: 66451678484b
The LDAP sync only ever granted the base "user" role. Add an optional
auth.ldap.role-filters map (role -> LDAP filter) so accounts matching a
filter are granted that elevated role (admin/support/api/manager).
LDAP is authoritative for the configured roles: sync both adds and removes
them to match group membership, while roles not listed (e.g. a manually
granted manager) are preserved. A managed manager that still has assigned
projects is never stripped. With no role-filters configured behaviour is
identical to before.
Roles are reconciled during periodic sync and at login. Sync evaluates
each filter once over the whole base (one search per role, not per user)
and reconciles existing users via a single ListUsers lookup plus the new
UserRepository.UpdateRoles helper.
Closes#74
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: e38526c3259d
Bump gqlgen, aws-sdk-go-v2, go-oidc, go-sqlite3, gqlparser, go-openapi,
golang.org/x/* and other transitive dependencies to their latest patch
and minor releases.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: 964ca6774882
The backend resample config changed from {trigger, resolutions[]} to a
policy/targetPoints model, but several components still spread
resampleConfig.resolutions, throwing "resolutions is not iterable" and
breaking the job list render. Default the initial resolution to null (the
backend resolves it from the configured policy; zoom overrides it) and
drop the now-obsolete resolution selector and admin display fields.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: 27516ce259fd
A setting nested at the wrong level (e.g. "resampling" placed next to
"main" instead of inside it) was silently ignored. Emit a startup warning
for any top-level config section not consumed by the backend so such
misconfigurations surface instead of failing silently.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: 69d1b85c8db3
Add a "main.footer-links" config option so the footer Imprint and
Privacy Policy links can point at internal pages (default) or external
URLs. External http(s) targets open in a new tab; empty/unset values
fall back to the built-in /imprint and /privacy routes, keeping the
existing ./var/*.tmpl override mechanism intact.
Closes#517
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Entire-Checkpoint: c5dca72c848f
moebiusbandandClaude Opus 5
Christoph Kluge
Aditya Ujeniya
Jan EitzingerandGitHub