Changelog¶
1.1.7 — 2026-11-13¶
- UI: fold state is kept in the browser. Clusters, namespaces and workloads the user folded or unfolded (one by one or with expand/collapse all) open the same way next time; a workload with more than 5 pods still starts folded the first time, but once unfolded it stays unfolded. The selected-streams bar keeps its ▸/▾ choice too.
- Fixed: data race on the agent's counters (
Shipper.Sent/Failed,Batcher.Dropped), written by the shipper and batcher goroutines and read by/statusand the heartbeat without synchronization; they are atomics now.Tailer.Key()is safe to call while the tailer runs. Both found by running the new unit tests under-race. - Fixed: the hub logs failed catalog updates, object-storage deletes and the WAL checkpoint on close instead of dropping the error (the index stays rebuildable; an unhealthy disk is now visible in the log). Chunk recovery treats a frame whose payload does not decode as corrupt and truncates there instead of indexing it. WAL open fails loudly on a segment file with an unexpected name instead of silently taking sequence 0. Export and tail stop writing once the client is gone. The agent logs a failing health listener and a failed checkpoint save at shutdown; the hub logs when the session key cannot be persisted.
- Tests: unit coverage for
index(0 → 91 %),federation(0 → 100 %),tailbus(0 → 100 %),hubconfig (0 → 100 %),agent(0 → 89 %, including a fullRunagainst a fake hub),api(5 → 93 %) andenrich(8 → 97 %);make testruns with-race. - CI:
golangci-lintv2 (.golangci.yml,make golangci), unit tests with-raceand coverage upload to Coveralls, badges in the README. Fixes for every lint finding (US spelling, unchecked errors, dead assignments, tagged switch, HTTP status constants). - Release: both images are signed with cosign (keyless) and carry an SPDX SBOM and SLSA
provenance (
mode=max) next to the signed chart; the GitHub release body carries the verify command and the image digests. README documents verification. - Bench:
bench/run.shreports lines written, lines ingested and spool drain time, and waits for the agent to drain before measuring disk. New numbers on a Linux VM (2 k / 10 k / 30 k lines/s) and 34 days of steady state on a 3-node k3s cluster inbench/RESULTS.md; README footprint table updated from them.
1.1.6 — 2026-11-09¶
- Security: images move to
gcr.io/distroless/static-debian13:nonroot(the Debian 12 base shipped ca-certificates DLA-4845-1 and tzdata DLA-4792-1, DLA-4854-1, DLA-4867-1) and are built with Go 1.26.10 (CVE-2026-111943, crypto/x509 distrust of compromised intermediates). go-jose 4.1.9 fixes CVE-2026-116585, CVE-2026-116586 and CVE-2026-116587 (hub, OIDC ID-token verification). klauspost/compress 1.20.3 fixes CVE-2026-116591, a zstd decoder panic on a crafted frame reachable through the hub's ingest endpoint. Clears the 13 findings Artifact Hub reported for 1.1.5.
1.1.5 — 2026-10-19¶
- UI: namespaces can be pinned with the star next to their name; the star button in the sidebar shows only pinned namespaces. Pins are kept in the browser.
- UI: A-/A+ (keys
-and+) set the log text size, 11–20 px; row height follows so virtual scrolling stays exact and the reading position is kept. - UI: system / light / dark theme switch in the header, applied before first paint.
- UI: log text 13 px, sidebar text +1 px, brighter timestamps and metadata (faint tone now ~5:1 contrast in both themes).
- Dev:
make devrebuilds and restarts the hub on Go edits (a failed build keeps the previous hub running) and, with-dev-ui, servesui/index.htmlfrom disk and live-reloads open browsers. Normal builds keep the embedded UI.
1.1.4 — 2026-10-07¶
- Release: chart releases are signed with cosign (keyless, GitHub OIDC).
1.1.3 — 2026-10-07¶
- Chart: Artifact Hub repository metadata (verified publisher), category
monitoring-logging, Artifact Hub badge in the README.
1.1.2 — 2026-10-06¶
- Hub, basic auth:
hub.auth.ui.basic.lockPassword=truefor demo hubs. The password from the chart secret becomes the only accepted credential: the users file on the data volume is not opened (so a password someone changed earlier no longer takes precedence),/auth/passwordanswers 403 and the UI hides the "password" link. The chart refuses the flag withouthub.auth.ui.basic.passwordorexistingSecret; the hub refuses it without a configured username and password.
1.1.1 — 2026-10-05¶
- UI: the bar listing the selected streams above the log view can be collapsed to a single line (▸/▾ button at its left, hover for the live/gone count). It folds by itself once more than 8 streams are selected, so tailing a whole CronJob no longer pushes the logs off the screen; the choice sticks for the session.
1.1.0 — 2026-10-04¶
Real-cluster readiness. No format or API changes.
- Multi-node, multi-cluster test environment on kind:
make e2e-multibuilds the images, creates a 3-node hub cluster (tainted control plane, hub as NodePort 30080 on the host) and a 1-node spoke cluster that pushes over the docker network with a per-cluster token, deploys the demo workloads to both, and checks DaemonSet placement on every node, cross-cluster queries, label enrichment, restart boundaries, CRI partial-line reassembly (20 KiB lines), multiline stack traces, rate-limit markers, token scoping, short-lived pods (CronJob, Job with init container) and a hub outage (spoke spools, then drains with no loss and no duplicates). It prints the UI URL and credentials. - Demo workloads in
hack/demo/(busybox only): JSON API with a sidecar, crash-looping worker, nginx-style frontend, 400 lines/s spammer, 20 KiB line writer, Spring-style app with stack traces, a CronJob and a Job with an init container, plus a two-container payments gateway.make demoapplies them to the current context. - Agent, multiline: groups now honour
multiline.timeout(they were flushed at every batch interval, so a trace written across an interval tick was split); a group is flushed as soon as its file disappears; size-triggered flushes now work in multiline mode (a batch could grow pastbatch.maxBytes/maxLinesuntil the next tick). Unit tests for the batcher (multiline join, timeout, stream end, size flush, rate-limit marker and checkpoint advance). - Agent, checkpoints: a file is now identified by path, inode and a hash of its first
line. Kubelet deletes a rotated file after gzipping it and the kernel reuses the freed
inode for the next
0.log; keyed by path and inode alone, the new file inherited the old checkpoint and the hub's per-file cursor, so every new line was dropped as a duplicate until the file outgrew the old one, which at 10 MiB rotation never happens. Observed on kind and on k3s after ~3 hours (one container silently stopped at its second rotation). The fingerprint covers the first line or the first 64 KiB, whichever is shorter (a containerd record can be 16 KiB before its first newline). Legacypath@inodecheckpoints are adopted once, so upgrading does not re-send files. In-place truncation (CRI-Olog_size_max) now also starts a new identity instead of reusing the key. - Agent, tailer: after kubelet renames
0.log, the runtime keeps writing to the old inode until it reopens the path. The tailer's own 2 s inode check used to stop reading as soon as it saw the path point at a new file, losing those lines about one rotation in eight (caught bymake e2estep 4 under load). It now drains the old inode for the rotation grace period, and ends a stream only when the path is still gone on a second check ~1 s later (the path is briefly absent between rename and reopen). Regression tests ininternal/tail. - Agent, shipping: a push now has a bounded deadline (10 s plus 1 s per MiB; dial 5 s, response header 10 s). Before, when the hub pod vanished while a keep-alive connection was open, the next push hung for the full 60 s client timeout before the agent started spooling, and every tailer on the node paused behind it (measured on kind: 64 s, now ~10 s). Found by the multi-cluster e2e's hub-outage check.
- Chart:
hub.service.nodePort(fixed NodePort for kind / dev clusters); the defaultagent.multiline.startPatternalso recognises klog (I1004), logfmt (time=) and access logs starting with an IPv4 address, so enabling multiline no longer glues system-component lines together. - UI, sidebar: pods are grouped under their workload (Deployment / StatefulSet / DaemonSet / Job / CronJob / Pod) with a kind badge, pod count, restart total and one click to tail every container of the workload; workloads with more than five pods (CronJob runs) start folded. Under the search box: unselect all (with count), expand all, collapse all, a "live" toggle that hides pods that no longer exist, a namespace selector and workload-kind chips. The search also matches workload names, kinds and labels.
- Agent, enrichment: when
agent.enrich.enabled, every stream carries a synthetic_ownerlabel (Deployment/api,CronJob/report, …) derived from the pod'sownerReferenceswithout extra API calls;label=_owner:Deployment/apiworks as a selector. Without enrichment the UI guesses the kind from the pod name (DaemonSet and Job pods are indistinguishable by name and are shown asds·job). - UI, log view: only visible rows are rendered. Fixed-height mode was already virtualised; wrap mode now keeps a window of 300 rows in the DOM and reveals earlier rows as you scroll up (first from memory, then one page from the hub). Rows are rendered once and cached; live batches append instead of rebuilding the whole view; the in-memory buffer is capped at 10 000 lines and older lines are re-fetched on scroll-up, so a pod logging hundreds of lines per second no longer makes the tab lag.
hack/incus-k3s.sh: create N Incus VMs on a host and join them into one k3s cluster (used for the first multi-node test outside docker).make images-tar ARCH=amd64cross-builds for the target nodes.- API: relative times accept days and weeks (
start=-7d,-2w,-1.5d). Go'sParseDurationrejectsd, so-7dsilently became the one-hour default: the UI's 7d preset returned fewer lines than 24h. Unit test forparseTime, e2e check with a two-day-old line. - Auth, first run:
ui.mode: basicis now a login form with a session cookie. Without a configured password the hub seedsadmin/adminand forces a new password on the first sign-in (stored as PBKDF2-SHA256 in/data/auth/users.json, so it survives restarts and upgrades); "password" in the header changes it later. A password fromhub.auth.ui.basic.password/existingSecretstill works and skips onboarding. HTTP Basic headers keep working for scripts. The session key is generated once under/data/auth/session.key, so sessions survive hub restarts. The chart no longer generates a random password secret by default. - UI: in fixed-height mode a message with embedded newlines (a joined stack trace) grew its row beyond the 20 px the virtual list assumes, so the scroll height drifted and the bottom could not be reached (jumping to it emptied the view and clamped the scroll back). Rows are now clamped to one line with newlines shown as ⏎; Wrap mode still shows them in full.
- UI: an empty result says when the newest line of the selection was and offers the range that covers it ("No lines in the last 1h · newest 1h 21m ago · Show 6h").
- UI: the address bar always equals the permalink (selection, filter, range, follow, wrap, JSON, timestamp mode, tab), so copying the URL works as well as the Permalink button; wrap / JSON / timestamp mode are restored from the URL on load.
- UI: a round ⤓ button in the bottom-right corner of the log view appears whenever you are scrolled up and jumps back to the newest line (also the End key); the "↓ N new lines" pill keeps counting what arrived meanwhile.
- UI: the "Pick a container on the left" overlay never went away because its
display:gridrule beat thehiddenattribute; a global[hidden]{display:none!important}fixes it. - UI: the Agents tab rendered its tiles and table inside the (hidden) stream-tree column
and left the main area empty; it now spans the full width. Switching tabs updates the
URL hash, so
#agentslinks and reloads land on the right tab. make images-tarsaves both images to one tarball for nodes without registry access (docker load/ctr -n k8s.io images import);make kind-up,make kind-down.make e2e: the object-storage step now waits until every chunk is sealed and copied before wiping the data dir, so the rebuild-from-bucket check is exact instead of timing-dependent; the test hub for that step is killed on exit; test hubs bind127.0.0.1explicitly (on some macOS hosts IPv4 connections to a dual-stack:portlistener hang in SYN_SENT once Docker networks exist).- README: screenshots and a live-tail GIF (
docs/img/), generated from a seeded local hub; the docs site rewrites the image links. - License: copyright notice filled in (2026 p10node),
NOTICEfile, OCI license/source labels on both images. - CI:
kind-multijob runs the multi-cluster e2e.
1.0.0 — 2026-10-01¶
Format and API freeze (see docs/FORMATS.md).
- Storage v2: write-ahead log + per-stream memtables. Pushes are fsynced to the WAL and
acknowledged, then written to chunks as large frames (≥ 256 KiB or 30 s), so quiet
pods no longer produce one tiny frame per second.
kill -9recovery replays the WAL with per-stream high-water marks: no loss, no duplicates. - Federation: a hub can front other hubs.
/streams,/query,/tail,/export,/statusfan out and merge by timestamp; peer stream ids are<peer>/<id>; loops are stopped byX-P10-Hops. - Object-storage offload: sealed chunks older than
uploadAfterare copied to any S3-compatible bucket (built-in SigV4 client, no SDK), local copies are evicted first when the disk cap is hit, queries fetch cold chunks transparently,--rebuild-indexlists the bucket. - Per-cluster ingest tokens (
auth.clusterTokens) and viewer roles (auth.roles: users / OIDC domains / API tokens scoped to cluster and namespace globs). p10logs-hub --checkverifies chunks, WAL and index.- Helm:
values.schema.json, Artifact Hub metadata, upgrade path tested on kind. - Docs site (mkdocs) and this changelog.
0.2.0 — 2026-10-01¶
- Label/annotation enrichment via one node-scoped pod watch (no client-go);
label=key:valueselectors. - Trigram bloom filter per chunk;
skipped_chunksin query stats. - Fixed: Helm rendered two different ingest tokens (agent 401 forever); spool now drains on a timer; heartbeat failures are logged.
- CI (unit, local e2e, chart, kind) and release (multi-arch images, OCI chart) workflows.
- Benchmarks in
bench/.
0.1.0 — 2026-09-30¶
- First working release: agent, hub, UI, Helm chart.