Changelog

All notable changes to dlboard are documented here. The format follows Keep a Changelog; versioning follows the tiers described in docs/compatibility.md.

[Unreleased]

[0.8.0] - 2026-10-16

Added

• A table chart’s rows are a setting you choose: a run, each value of a metric (step, epoch, …), or each logged artifact. With rows of a metric, spread_by_run puts the runs side by side, so “a row per epoch, each run’s latest image of that epoch” is one table. What a run logged between two values of the metric belongs to the earlier one. A metric can also be reduced to its first, minimum, maximum or mean (aggregate) instead of its latest. See docs/plugins/charts.md.
• A table’s metrics, hyperparameters and artifacts take names or a prefix and * (eval/*, or * for all). In the chart editor you can type a prefix instead of picking from the keys logged so far, and a prefix also matches keys logged later.
• Histograms. DLBoardLogger.log_histograms({"grad_norm/layer1": tensor}, step=...) logs the distribution of any array, tensor or list of numbers at a step – per-layer gradient norms, weights, per-sample losses – instead of averaging it into a scalar. The values are binned in your training process, at the call, so only bucket edges and counts (at most DLBOARD_MAX_HISTOGRAM_BUCKETS, 1024 by default) are sent and a bad binning slows down only the script that chose it. ExponentialBinning (the default) puts a fixed number of buckets per power of ten on a lattice that does not depend on the data, so steps and runs line up bucket for bucket; ExplicitBinning takes the edges you give. See docs/client.md.
• A histogram chart shows them as a heatmap over steps, as percentile bands, or one step at a time with a slider, a row per run in that run’s color. A chart reads about DLBOARD_MAX_HISTOGRAM_STEPS (256) steps per run however long it trained, thinned in the database by a power-of-two stride, so it costs the same to open after a million steps as after a thousand. Its colorscale, the share of values its bands hold, and how many rows and slider steps it draws are chart settings, and its accent color and spacing are DLBOARD_HISTOGRAM_* settings (docs/configuration.md). Auto-generate and Suggest charts make a heatmap for each histogram key. See docs/plugins/charts.md.
• A database migration (0003) adds the histogram table. A histogram’s bucket edges and counts are stored as packed binary arrays (8 bytes an edge, and a count in as few bytes as it fits), not as JSON text: about a third of the space on SQLite, and none of the cost of printing and parsing every float. A server older than this ignores the histograms a newer client sends, so upgrade the server first.

Changed

• A table chart’s mode (runs, pivot, artifacts) is replaced by rows, row_field, spread_by_run, show_metrics, show_hparams and show_artifacts. A table saved with the old settings opens as the new ones that show the same table, in the chart editor too.
• Everything the server prints now goes through one logging setup. Granian’s own lines, and any library that uses the standard logging (Flask, Werkzeug, SQLAlchemy, …), are rendered like dlboard’s own, so DLBOARD_LOG_FORMAT=json is JSON throughout, DLBOARD_LOG_LEVEL governs Granian’s too, and DLBOARD_TELEMETRY_LOGS forwards them. A library’s records appear from WARNING up.

Fixed

• A table chart can show no hyperparameters (or no metrics): an empty list used to mean “all”, so choosing only an artifact brought every hyperparameter and metric along. New tables still start with all of both.
• An error the server didn’t handle is logged where you will see it. It used to be left to Flask’s own logger: unstructured, with no request id, and absent from JSON and OTLP logs. It is now one event, with the method, path and request_id, and the client’s error for that request names the same id (server request id ...), so a failed batch points at its log line.
• Starting the server no longer prints RuntimeWarning: Repeated configuration attempted once per worker.

[0.7.0] - 2026-10-14

Added

• A line or bar chart is titled with its full metric name, above the chart. The y-axis title is still shortened to fit, so a long name was hard to read.
• A table chart can show artifacts. In its new artifacts mode it lists one row per logged artifact under a key, or every key under a prefix (eval/, including keys logged later), with each tag as a column that sorts and filters as the number or text it is – images as thumbnails, other files as download links. In runs mode its artifacts setting adds a column per key with each run’s latest artifact. A tag measured in a unit is shown in its base unit (seconds, bytes) in a table, so runs compare. See docs/plugins/charts.md.
• A table chart’s mode (runs or pivot, besides artifacts) is now a setting of its own instead of being implied by pivot_on; a table saved with pivot_on and no mode still loads as a pivot, and a pivot without pivot_on and pivot_metric is a form error in the chart editor.
• A table chart names and colors its runs as the charts and the Runs table do (a pinned Run column, run names as pivot headers), groups its columns, filters in place, and keeps its sort, filters and page when a live update redraws it (it used to start over every time).

Fixed

• A table chart showed each run’s last row rather than each metric’s own latest value, so a metric logged less often than the last step (a per-epoch val_acc beside a per-step loss) was blank; it now agrees with the Runs table. It also no longer shows step as a metric, shows only hyperparameters when only those are chosen, keeps an hparam and a metric of the same name apart, and reads a metric named a.b as one column.
• Numbers in tables: whole numbers are shown as they are (not 12.000) and small ones to four significant digits (a learning rate of 1e-4 was 0.000).
• A live update redrew charts it should not have (another run logging a step renumbered the rows a chart was compared by) and missed ones it should have (a hyperparameter edited, a run renamed).

Changed

• Which panels are open and which tab is shown are part of the page’s address (?open=, ?tab=), like the page of panels already was, instead of being saved on the experiment’s page or view. Opening a panel used to open it for everyone else looking at the same page; now a link you share shows what you were looking at, and a fresh visit opens the first panel of the first tab. Pages saved before still load – the open panels and tab they recorded are just no longer read.
• Changing anything about the shared page, its settings included (the runs it leaves out, its Runs-table columns), now takes an editor, as changing its panels already did. A viewer’s edit still branches into a view of their own.

[0.6.0] 2026-10-13

Added

• A panel’s Sync crosshair and step across charts setting now also locks its image-series sliders together: moving one moves the others to the latest step each has at or before it.
• An image-series chart is titled with its artifact key.
• A tag can carry a unit: tags={"took": Quantity(value=5400, unit=TimeUnit.SECONDS)} records what the number was measured in (time ns to d, size B to TB, decimal); on the wire it is {"value": 5400, "unit": "s"}. The client never says how to show it: the image chart shows it in the largest unit that fits by default (1.500 h, 2.500 GB, not 5400 or 2500000000), or in the unit set by its new tag_time_unit / tag_size_unit settings, to tag_precision significant digits, and tag_display overrides those per tag name. Plain numbers are written with separators (12,346) instead of 1.235e+04.
• Opt-in OpenTelemetry for enterprise deployments. DLBOARD_TELEMETRY_METRICS=prometheus,otlp exports the server’s own metrics (HTTP requests and latency, metric and artifact ingest, the artifact write queue, run lifecycle, sign-ins, artifact purges) either as a Prometheus scrape of /metrics – added up across every worker – or pushed over OTLP to a collector, Mimir or Grafana Cloud, and DLBOARD_TELEMETRY_LOGS sends the logs over OTLP too (Loki). It is the new otel extra (OpenTelemetry SDK 1.25 or later, 1.38 for logs), installed in the Docker image; with none of it set nothing is imported or logged, and every process (each Granian worker and its artifact writer) sets itself up on first use, so a collector that is down never stops the server from starting. Telemetry never takes down what it measures: a failure to record is logged once and skipped, and a setup that failed is retried with a back-off. The metrics API is public, as dlboard.serve.telemetry, for plugins to record their own. See docs/telemetry.md.
• Docker images between releases: ghcr.io/dlangerm/dlboard:beta follows main (and sha-<commit> pins a single landing), :latest is the same tested main, and :stable follows the newest release. See docs/docker.md.

Fixed

• Live updates replaced a chart’s whole item with a whole new item nested inside it (a second set of controls and padding) the first time its data moved, and redrew every open chart on the first update after a page loaded, including ones whose data had not changed. A chart now carries a fingerprint of the data it was drawn from, so a poll replaces only the chart inside its item, and only when its data moved.
• A request the sign-in gate turned away (a 401, or the redirect to the sign-in page) had no request id, and with DLBOARD_ACCESS_LOG on its response raised an error instead of being logged.
• Pressing Ctrl+C during training left the process hung for a minute per unshipped queue, with “still hadn’t all been shipped” and “shipping process is dead” warnings. The terminal’s Ctrl+C reached the logger’s background shipping processes too, so they died before the final flush could send what was queued. They now ignore Ctrl+C, and the training process flushes and exits.

Changed

• The latest Docker tag no longer means “the newest release”: it now follows main (the same as :beta). Pin :stable (or a version, :0.5.0) for releases only.
• An image-series chart lays its runs out by what its panel and page say: as many columns as the panel’s grid column count (following it the instant it changes) and as many runs per page as the viewer’s charts-per-page size, instead of a fixed three and six. It never uses more columns than it has runs, so one run’s image fills the chart’s width and a large chart size now enlarges it.
• Artifact tags values are JSON scalars (str | int | float | bool) on the wire and in storage, not text (#179): NewArtifact.tags is dict[str, TagValue], clients send what the caller logged (tags={"score": 0.14}), and the image-series and file-list charts format numbers the same way wherever they come from. A server older than this release rejects a numeric tag with a 400, so upgrade the server before the clients; existing string tags are unaffected. This replaces the text-only storage described under 0.5.0.

[0.5.0] - 2026-10-11

Added

• The experiment page pages its panels and charts, so its cost no longer grows with the experiment (#168: 886 charts took about 30 s). It shows one page of the active tab’s panels (default 10) with a pager, and a fuzzy filter on panel names (vl finds val/loss); inside a panel, one page of its charts (default 12), with a filter on the metric and artifact names they show. Only the charts on screen are fetched, on load and on every live-update poll. The page and filters are in the URL (?panels_page=, ?panels_q=, ?charts=), so a link reproduces what you were looking at, and a ?chart= link lands on the page holding that chart. How many panels and charts per page is each viewer’s own choice, kept in a dlboard_page_sizes cookie.
• A page left open across an upgrade offers a Reload prompt instead of failing every live-update poll: the server answers a callback request in an older shape with a 409 (not Dash’s bare 500), and the page shows the prompt once.
• A run’s checkpoints are one list, not one chart per file. The file list takes a key_prefix (checkpoints/) in place of a key and shows every key under it, including ones logged later; Auto-generate and Suggest charts make exactly that for a directory of files, so a run that kept hundreds of checkpoints no longer gets hundreds of charts. The list is paged and sortable, and splits into Latest, Best and Top-k tabs from the tags DLBoardLogger records (any other file kind gets its own tab). A view with a key_prefix chart is not valid on an older server, so it can’t be opened after a downgrade; the height setting no longer does anything (the list pages instead of scrolling) but is still accepted.

Changed

• Only the active tab renders; the other tabs’ panels are no longer fetched and drawn behind it. Dragging a panel or chart to reorder it now rebuilds the page of panels (or charts) being shown, rather than moving the nodes already on screen, so it can only be dropped among what is showing. Opening and closing a panel keeps the open state of the panels on other pages and tabs.
• Line charts render much faster on a panel with many charts: the server no longer copies and scans the whole panel’s data once per chart (200 charts took 99 s of server time, now 4.6 s).
• Chart axes are readable: the value axis has room for its ticks beside the rotated title, ticks show at most four significant digits (no more 0.30000000000000004), and a line chart’s tooltip is titled with the whole metric name, which the axis title may shorten.
• The panel menu button is a sliders icon rather than three dots (which read as one more drag handle next to the grip), the header is padded off the panel’s right border, and each button that rebuilds the panel area (create, rename, delete) spins until it finishes.
• dlboard now depends on dlboard-client~=X.Y.0, a client of its own minor version, instead of any version, so an environment holding both can no longer end up with a mismatched pair.
• DataStore.fetch_artifacts takes a key_prefixes keyword (default none) that also selects every artifact whose key starts with one of the prefixes, and ChartType.hint_required_artifact_key_prefixes lets a chart ask for them. A custom DataStore must accept the new keyword; a custom ChartType needs no change.

[0.4.0] - 2026-10-09

Added

• A run records how it ended and when. DLBoardLogger.finalize reports "success" as finished and "failed" as failed (POST /api/v1/runs/<id>/finish, a new path), stamped with the client’s clock; the runs table gets a compact Status column (running, 3m 23s, 2d 3h, ✕ 1h 2m) that sorts by duration and filters by status. A script that exits without finalizing the logger is recorded as unknown; a process killed outright never reports, so its run stays running. Two nullable columns are added to Run by migration 0002 (expand-only: run dlboard servers upgrade first, as always); an older server answers 404 to the new path and the client carries on.
• Suggest charts is a selection list, not one add button per key: a filter (by key or target panel), a By name / By panel sort, Select all and Clear for what the filter shows, and Add selected (N), which adds everything ticked in a single page update (65 charts at once on the advanced example, instead of 65 clicks and 65 rebuilds). The one-click add on each row is still there.
• Artifact tags (on Image, Figure, File, Link and anything implementing AnyArtifact) accept numbers and booleans as well as strings, and format them identically wherever they come from (floats to four significant digits). Callers no longer stringify, and so no longer disagree. Tags are still stored as text – a numeric tag would be rejected by an older server, which the supported version skew allows.
• DLBoardLogger flattens nested hyperparameters, like Lightning’s own loggers do: a pydantic model, a pydantic or stdlib dataclass, a Namespace or a plain dict, nested however deeply, is logged as parent/child keys instead of being rejected. Pydantic does the serializing, so a model’s own aliases and field serializers apply. Lists are stored as a JSON string. See flatten_hparams.
• DLBoardLogger(log_model=...) uploads ModelCheckpoint checkpoints, like the MLflow and W&B loggers: "all" as each one is saved, True only the ones kept once training finishes. Backed by the new dlboard.client.artifacts.file.File artifact kind, which uploads a file in place and tags what it is (FileKind.CHECKPOINT).
• examples/lightning_advanced.py: a 25-epoch UNet on synthetic shapes with nested hyperparameters, Lightning’s learning-rate, throughput and device-stats callbacks, checkpoints, and segmentation and bounding-box previews – a deliberately heavy run for exercising the UI.

Changed

• A checkpoint (anything logged as a File) now gets a file list with download links when charts are auto-generated or suggested, instead of an image chart full of broken images. It’s the new built-in files chart type. Chart types now say which artifacts they can display (ChartType.can_display_artifact), and an artifact no chart type claims gets no generated chart at all.
• Image artifacts are encoded as lossless PNG by default instead of always as lossy JPEG, so masks, label images and overlays are no longer degraded on upload. Pass format="jpg" to keep the old, smaller files. Artifacts already stored are unchanged.
• A panel’s header shows how many charts it holds, so you can tell how big it is before expanding it.

Fixed

• dlboard and dlboard-client now import together when each is installed into its own directory on sys.path, as hermetic builds such as Bazel do. The client’s regular dlboard package used to hide the server’s half (dlboard.serve, dlboard.plugins), so depending on dlboard failed to import it. dlboard/__init__.py now extends its __path__ with pkgutil.extend_path. Ordinary single-directory installs are unaffected.
• Creating charts with Auto-generate no longer makes the page wait for the first panel to draw when that panel is big. A panel opens by itself only if it holds 12 charts or fewer (otherwise the first small one does, or none); a big panel stays closed with its chart count on its header. On a run with 886 charts the default grouping now shows its panels in about 1.6 s instead of 5.8 s.
• Switching Prefix/Suffix in the Auto-generate charts dialog no longer freezes the page for seconds on an experiment with many metrics: the preview lists the first 20 panels and counts the rest (239 panels took the browser about 3 s to draw; it takes about 0.1 s now).
• Changing a Grid panel’s charts-per-row no longer rebuilds every chart in the panel. The browser applies the new column count itself and the server only saves it, so a panel with over a hundred charts freezes the page for about a quarter of the time (0.8s instead of 3s per click, measured on 126 charts).
• A chart could show the wrong image when the browser had visited another dlboard database at the same address (a wiped or restored database, or a different --sqlite-location). Artifact URLs are keyed by id, ids start over in a new database, and the browser was told to cache them for a year; so artifact 5 of the old database was shown in place of artifact 5 of the new one. Artifact URLs now carry a version token (?v=) that changes with the artifact behind the id.
• A long metric name no longer spills its y-axis title out of the chart (over the chart header and the charts beside it): titles are shortened to fit, keeping the end of the name. How long a title may be is a setting of the line and bar charts (max_axis_label_chars, 36 by default).
• Metrics and artifacts logged by a later stage were never shipped when a script exited right after it – notably everything trainer.test logs after trainer.fit. Lightning finalizes the logger after every stage, but only the first call flushed.

[0.3.1] - 2026-10-08

• Reduce numpy requirement floor to 1.26 to expand compatibility.

[0.3.0] - 2026-10-08

Added

• DLBoardLogger.log_figure(figure, key, step), a near drop-in for mlflow.log_figure: renders a matplotlib figure (a lossless PNG by default, or whatever supported extension the key ends in) and logs it as an image-series artifact. Backed by the new dlboard.client.artifacts.figure.Figure artifact kind.
• Run compare: a Compare button above the experiment’s run table opens a modal that diffs up to five runs’ hyperparameters and latest metric values, as differences only or every key, with a search box over keys and values. Differing cells are highlighted against the first run picked. The picked runs, mode and search are kept in the URL, so copying the address (or the modal’s link button) shares exactly what you’re looking at.
• DLBOARD_SESSION_CLOCK_SKEW_SECONDS (default 120): how far ahead of a replica’s own clock a session cookie may be dated and still be accepted. See docs/configuration.md.

Fixed

• People are no longer randomly signed out, or bounced with a 400 from the login form, when a session cookie was signed by a replica (or a clock) a moment ahead of the one reading it. Flask rejected such a cookie outright and treated the browser as signed out; under several replicas, or across a clock step, that hit perfectly good sessions. Only the lower bound is relaxed – a cookie older than the session lifetime still expires as before, and revoking a session is unaffected.

[0.2.1] - 2026-10-07

• Lower version requirements to enable broader install base
• Fix Dash in tab title
• Adjust readme

[0.2.0] - 2026-10-07

• Change dlboard to install the full fat server+client and dlboard-client to install only the pytorch lightning logger to make it more straightforward to install for tensorboard-like use cases.

[0.1.3] - 2026-10-07

Added

• First public release candidate
  • Builtin charts: bar chart, line chart, table, image series
  • Custom plugins for extensible charts, backend functionality, logging, theme, and more, all powered by Dash/Plotly
  • Asynchronous hyperparameter and metric logging
  • Pytorch lightning logger support out of the box
  • Local serve/client experience like tensorboard
  • Production ready, s3 and postgres backing file and data stores built in