Repository navigation
air: Report each --watch STATUS transition once and never after the run ends - #7621
Merged
Merged
Conversation
…un ends When the Bricklens stream of `air run --watch -o json` hands off to the MLflow fallback (an empty terminal stream, FEATURE_DISABLED, or MLflow logs found while active), the fallback restarted transition tracking from the status resolved before streaming. A run cancelled while PENDING therefore emitted PENDING, CANCELED, PENDING, CANCELED; a FEATURE_DISABLED handoff emitted PENDING twice; an active handoff regressed RUNNING to PENDING. - The STATUS sink in `air run --watch -o json` drops repeats and anything after a terminal state, and reports the last state it actually emitted as previous_status. - fetchLogs passes the status Bricklens last observed into the MLflow fallback, which reports it instead of the stale pre-stream status. The fallback still follows on the pre-stream status, so a followed run keeps the active MLflow stream (lenient first pass, grace wait, re-read, uncapped output, partial-line hold-back) rather than the capped completed tail. Co-authored-by: Isaac <no-reply@databricks.com>
Collaborator
✅ Integration test reportCommit: e79190e
Top 15 slowest tests (at least 2 minutes):
|
Co-authored-by: Isaac <no-reply@databricks.com>
vinchenzo-db
approved these changes
Nov 16, 2026
Comment on lines
+89
to
+94
| // mlflowLogFallbackReporting is mlflowLogFallback for a handoff from a | ||
| // Bricklens stream that may have seen newer state (latest) than status. The | ||
| // path and loop still follow status: a run that was being followed keeps the | ||
| // active stream, whose lenient first pass, grace wait and re-read let late | ||
| // artifacts print uncapped, instead of switching to the capped completed tail. | ||
| // Only the reported lifecycle state comes from latest. |
Contributor
There was a problem hiding this comment.
Can boil this down to
// mlflowLogFallbackReporting reports the latest Bricklens status while preserving active-stream behavior to drain late logs uncapped.
Co-authored-by: Isaac <no-reply@databricks.com>
Collaborator
❌ Integration test reportCommit: 12cd819
40 interesting tests: 24 FAIL, 10 RECOVERED, 3 KNOWN, 3 flaky
Top 50 slowest tests (at least 2 minutes):
|
deco-sdk-tagging Bot
added a commit
that referenced
this pull request
Nov 19, 2026
## Release v1.27.0 ### Notable Changes * `databricks aitools install`, `update` and `uninstall` now exit with code 1 when they do nothing they were asked to. `install` fails when it sets up nothing: no coding agent is detected, every agent is skipped or fails to install, or no skill in the release can be installed (also with `--path`). `update` fails when every update or check it attempts fails, and `uninstall` when every removal it attempts fails. Scripts that ran these commands as a best-effort step and relied on exit code 0 should handle the new exit code. ([#7481](#7481)) * Add `file:<path>` selectors to `pipelines run` and pipeline `bundle run`. `--refresh` and `--full-refresh` now use space-separated lists, such as `--refresh "orders payments"`, instead of comma-separated lists. Single names containing spaces now require backticks, e.g. ``--refresh 'table:`order details`'``. ([#7184](#7184)) * The Databricks CLI is now available on PyPI as `databricks`: run it with `uvx databricks`, or install it with `uv tool install databricks`, `pipx install databricks` or `pip install databricks`. The package contains only the prebuilt CLI binary. Uninstall the legacy `databricks-cli` package first, since both install a `databricks` command. ([#7704](#7704), [#7777](#7777)) ### CLI * Add `databricks sandbox fs` commands (`cp`, `ls`, `cat`, `mkdir`, `rm`) to transfer files between your local filesystem and sandboxes using `<sandbox-id>:/path`. ([#7372](#7372)) * `databricks fs` commands retry reads (stat, list, open) and deletes on DBFS and Volumes when the server returns a transient error or does not respond within the HTTP timeout. ([#7556](#7556)) * Retry UC Volumes file operations (`fs cp`, `fs ls`, `fs rm`, bundle uploads to Volumes) when the Files API returns HTTP 408, 429, 502, 503 or 504 (for example "AWS S3 is throttling requests"), or on a transient network error; this regressed in v1.10.0. ([#7559](#7559)) * Stop warning when the host metadata request (`/.well-known/databricks-config`) times out; the CLI already falls back to your config, and the message is still shown with `--debug`. ([#7752](#7752)) * Add `databricks aitools auto-install disable` and `enable` to turn off (or back on) the automatic installation of Databricks AI tools on CLI upgrades. A global-scope `databricks aitools uninstall` also turns it off. ([#7701](#7701)) * Stop printing the Databricks AI tooling install hint on `auth`, `configure`, `current-user me`, `version`, `help`, and shell-completion commands. ([#7452](#7452)) * When Codex, Cursor, GitHub Copilot CLI, OpenCode, Gemini CLI, Antigravity, Pi, Goose or Kiro run the CLI without Databricks skills installed for that agent, the CLI now prints a one-line hint to install them with `databricks aitools install --agents <agent> --scope global`, previously shown only to Claude Code. ([#7586](#7586)) * `databricks aitools install` now rejects an unknown `--agents` name before it shows the scope prompt, instead of after. ([#7469](#7469)) * `databricks aitools install` no longer asks "Proceed?" after the agent picker; the picker selection confirms the plan. The picker title now shows the install scope, its hint says Enter installs, `--skills-only` runs label plugin agents "skills" and pre-check every agent whose config dir exists, and the plan summary after it is reduced to the scope line. ([#7478](#7478)) * `databricks aitools install --output json` reports more specific `error_category` values: `SKILLS_FETCH_FAILED`, `RELEASE_LOOKUP_FAILED`, `RATE_LIMITED`, `MANIFEST_INVALID`, `STATE_INVALID` and `PLUGIN_CLI_TIMEOUT`. A coding agent CLI that does not answer in time is now reported as `timed out after <duration>`, followed by any stderr it wrote before it was stopped. ([#7475](#7475)) * `databricks aitools install` no longer prompts when a coding agent runs it, even in a pseudo-terminal; it uses the non-interactive defaults instead and prints a note saying so. ([#7477](#7477)) * `databricks aitools install` now prints the directory it installed skills into (for example `Installed 1 skill in ~/.databricks/aitools/skills.`), so agents can open the installed `SKILL.md` files without searching for them. ([#7668](#7668)) * `databricks aitools install` now ends a failed install with a step to recover, in the text output and in a new `recovery` field of `--output json`. ([#7480](#7480)) * `databricks aitools install` now installs plugins for the other coding agents even when installing skill files fails. ([#7479](#7479)) * `databricks aitools install` now installs skill files for a detected coding agent whose CLI is not on PATH (for example a Claude Code desktop or IDE install), unless it already has the plugin, says how to switch to the plugin later, and warns when the current project also has the plugin. `databricks aitools update` no longer tries to install the plugin for such an agent, and no longer writes skill files next to a plugin that such an agent already has. ([#7482](#7482)) * Fix `databricks aitools install` waiting a long time on an unreachable or slow workspace while it resolves authentication for telemetry; that wait is now limited to a few seconds. ([#7468](#7468)) * `databricks aitools install`, `update` and `list` now refuse a skills manifest that names a skill or file outside the skills directory, before writing any skill files. A skill whose name or file paths are not portable is skipped with a warning, and `update` keeps its installed copy. ([#7722](#7722)) * `databricks aitools install` and `databricks aitools update` refresh only the marketplace the Databricks plugin is listed in, so another unreachable marketplace no longer blocks them. When that refresh fails, both commands warn, install or update the plugin from the local copy of the marketplace, and record the plugin version as unknown. ([#7473](#7473)) * `databricks aitools install` and `databricks aitools update` now wait up to 20 seconds instead of 5 for a coding agent's CLI to answer the plugin capability check, so a slow first start is less likely to report the agent as unavailable. ([#7472](#7472)) * `databricks aitools install` no longer fails a plugin-only install when the lookup of the latest skills release on GitHub fails or is rate-limited. The plugin is installed and its version is recorded as unknown. ([#7474](#7474)) * `databricks aitools` commands now send usage telemetry, and the CLI records when it recommends installing Databricks AI tools. When the configured workspace is slow or unreachable, `aitools install`, `update` and `uninstall` can take up to about 6 seconds longer, and `list`, `version` and runs that stop early up to about 3 seconds. Set `DATABRICKS_CLI_DISABLE_TELEMETRY=1` to opt out. ([#7467](#7467), [#7469](#7469), [#7470](#7470)) * `databricks aitools install` and `databricks aitools update` now report a coding agent CLI shim (`.cmd` or `.bat`) on Windows that prints its plugin help and then exits with an error as `shim-exited-with-error`, with the command to run to see the error, instead of reporting the CLI as not on PATH. ([#7476](#7476)) * `apps init` no longer suggests installing skills when every detected coding agent already has the Databricks plugin. ([#7467](#7467)) * `databricks auth login` now gives up on the optional workspace lookup after about 5 seconds instead of retrying for about 5 minutes, and explains the fallback to the manual workspace ID prompt. ([#7522](#7522)) * Throttle Claude Code's native plugin install hint separately from the text hint: it prints at most once per 10 minutes per session, so the install dialog still comes up when a subagent, a hook, or a command outside a Claude Code subprocess runs the session's first `databricks` command. ([#7692](#7692)) * Print Claude Code's native plugin install hint alongside the text hint, so Claude Code offers a one-click install of the Databricks plugin. Claude Code now counts as having the tooling only when it loads it: skills only in `~/.databricks/aitools/skills` don't count, and a project- or local-scope plugin counts only when the CLI runs inside that project. `auth token`, `auth docker token` and CLI subprocesses no longer print install hints. ([#7585](#7585)) * Interactive prompts and pagers now work in Git for Windows' bash and other MSYS2 shells when they run under ConPTY (the default since Git for Windows 2.47) or directly in a Windows console. ([#7699](#7699)) * Add the `databricks feedback` command to send feedback about the CLI to Databricks. ([#7724](#7724)) * Fix shell completion in `databricks fs` commands for local paths typed with a leading `./` or repeated separators, which previously offered no suggestions. ([#7679](#7679)) * Fix a panic in `databricks labs` commands when a non-string SDK setting such as `DATABRICKS_DEBUG_HEADERS` or `DATABRICKS_RATE_LIMIT` is set in the environment. ([#7673](#7673)) * Fix a panic in `databricks labs install` for projects whose `labs.yml` has no `install` section. ([#7673](#7673)) * `NO_COLOR` no longer disables prompts, spinners or the list pager; they run without color. ([#7700](#7700)) * Add the Beta `databricks pipelines test` command, which runs the tests of a pipeline, and the `testing` property of pipelines in bundles. ([#7429](#7429)) * Add `databricks sandbox forward`, which serves an app running in a sandbox at a local URL. ([#7395](#7395)) * Bound the whole telemetry upload at exit, including config resolution and authentication, by its timeout, so a stalled auth or telemetry endpoint no longer keeps a command from exiting. ([#7467](#7467)) * `databricks version --check` now suggests `databricks aitools install` on stderr when a coding agent is detected without the Databricks plugin or skills. ([#7471](#7471)) ### AI Runtime * `air get -o json` now reports `PENDING` while the run waits for compute, matching `air list`, and includes the run's `status_message`. ([#7455](#7455)) * Wrap long failure reasons and log lines in the `air list` info and logs panes, keep the MLflow link (labeled with its retry) when a run's last retry failed before starting, and point `air logs` at earlier retries when the last one has no logs. ([#7456](#7456)) * `databricks air logs --node` now reports an out-of-range node index when it can determine the run's node count, instead of looking like a successful run with no logs. ([#7516](#7516)) * Add experiment, duration, accelerator, and dashboard metadata to `air list --output json`. For retried runs, `started_at` now reports the latest attempt, matching `air get`. ([#7568](#7568)) * `air list` shortens MLflow run names and unnamed-run IDs by keeping their end behind a leading `…`, and its columns no longer overflow on wide (for example CJK) text; `air get` shows an unnamed run's full MLflow run ID. ([#7582](#7582)) * Shrink the Experiment and User columns of the interactive `air list` table to fit the terminal width, keeping the end of a shortened experiment name. ([#7583](#7583)) * Show the experiment, accelerators and configuration in the `air list` info pane for runs where it showed `N/A`. ([#7587](#7587)) * Strip terminal control characters from the MLflow run, experiment and user names shown by `air list` and `air get`, so a run name cannot clear or retitle the terminal of whoever lists it. ([#7601](#7601)) * `air logs`, `air get`, and `air run --watch` no longer make an extra credential-check request before fetching the run, and `air get` reuses the configured workspace ID instead of looking it up. ([#7613](#7613)) * `air list` no longer waits on extra current-user requests before listing runs, which makes the default list about 150ms faster. ([#7615](#7615)) * `air list --all-status` requests larger index pages, so it makes far fewer round trips once the server supports `page_size`. ([#7616](#7616)) * Fit the interactive `air list` table to narrow terminals: when the Experiment and User columns are at their floors, shorten Started to `MM-DD HH:MM` and size the MLflow column to its labels, then hide Accelerators, User, MLflow and Progress in that order, naming the hidden columns in the hint line. ([#7629](#7629)) * `air get -o json` now includes `mlflow_run_id` and `mlflow_experiment_id`, so a run can be handed to MLflow-based steps without parsing `mlflow_url`. They identify the MLflow run `mlflow_url` links to, including one from an earlier retry that `mlflow_retry` names. ([#7636](#7636)) * `air get` and the `air list` info pane now show the environment of runs on the Databricks AI environment (`databricks_ai_v*`) instead of `N/A`. ([#7637](#7637)) * `air run` now rejects a `code_source.include_paths` path that selects no files instead of uploading a snapshot without it; glob patterns and pathspec magic are left to git. ([#7643](#7643)) * `air convert-to-dabs` now suggests the `--output-dir` that converts a working-tree workload whose `code_source.root_path` is outside the bundle directory, instead of suggesting `sync.paths`. ([#7644](#7644)) * `air run` now lists a `code_source.root_path` outside a git repository relative to `root_path`, so running from a subdirectory of `root_path` no longer uploads an empty or partial snapshot. ([#7646](#7646)) * `air list --all-status` now fills `--limit` past purged or inaccessible indexed runs. ([#7743](#7743)) * `air run` now warns about missing PyTorch when a pinned snapshot's `include_paths` selects a Python file. ([#7744](#7744)) * `air cancel JOB_RUN_ID` now rejects run IDs that are not AI Runtime runs instead of cancelling unrelated Jobs runs. In `-o json` output, the `failed[].error` text for a run that does not exist or that you cannot view now starts with `failed to get run <id>:` instead of the cancel API error. ([#7569](#7569)) * `air get -o json` now reports `mlflow_retry` when the MLflow link comes from an earlier retry because the run's last retry never reached MLflow. ([#7620](#7620)) * `databricks air run --help` now shows how to install the databricks-ai-runtime coding agent skill. ([#7652](#7652)) * `air run --watch -o json` now reports each lifecycle `STATUS` transition once and never reports a non-terminal state after the run has ended, including when logs are served from MLflow artifacts. ([#7621](#7621)) * `air convert-to-dabs` now explains the `--output-dir` workaround when `code_source.root_path` resolves to the bundle root, and notes when the generated job carries the default `max_retries`. ([#7579](#7579)) * `air run` now refuses to upload a `code_source` snapshot that selects no files (for example when `root_path` is git-ignored) and explains the cause, and `--dry-run` runs the snapshot file selection and reports the file count. ([#7570](#7570)) * Repeat the Jobs run link after live log streaming ends. ([#7571](#7571)) * `air get -o json` includes run details including flat `accelerator_type` and `num_accelerators`. Use `--include-config` to include stored config, which may contain plaintext `env_variables` and parameters; downloads over 1 MiB are omitted, and config-dependent `max_retries` is otherwise null. ([#7624](#7624)) * `air get` and the `air list` info pane no longer show `Max Retries 0` when the Jobs API leaves the task's retry limit out. For a run submitted with `air run`, the limit now comes from the run's uploaded config, shown as `3 (CLI default)` when the config doesn't set `max_retries`; it shows `N/A` when the limit can't be known (job runs, multi-container runs, legacy runs with no explicit value, or an unreadable config). ([#7623](#7623)) * Reorganize `databricks air` help: subcommands are grouped by task, `air run` has examples, and YAML fields are listed under headings, with pool fields marked optional and `command`, `compute.num_accelerators`, and `compute.accelerator_type` correctly marked required. ([#7512](#7512)) * `air run -h config` now lists the environment variables AI Runtime sets on every node (such as `MLFLOW_RUN_ID` and `CODE_SOURCE_PATH`), and `air logs --help` notes that `--download-to` cannot be combined with `--tail` or `--minutes`. ([#7655](#7655)) * Support full workspace accelerator types in `air list` filters while preserving family matching, and reject invalid filter values. ([#7515](#7515)) * `databricks air list --all-status` now warns when the AI Runtime run index is unavailable and it falls back to scanning Jobs runs. That fallback scan is capped at 250 runs, and a stuck index request fails after 15 seconds instead of retrying for minutes. `--all-status --filter user=<your email>` now uses the run index instead of scanning the workspace's full run history. ([#7619](#7619)) * Return structured JSON errors for invalid `air list --filter` values when `-o json` is used. ([#7612](#7612)) * `air list --all-status` no longer fails when a listed run has been purged. ([#7523](#7523)) * **Behavior change:** `databricks air logs` now prints the logs available now and exits, also for an active run, matching `kubectl logs`, `docker logs` and `databricks apps logs`. Pass `--follow` (`-f`) to stream new lines until the run completes; `--tail N --follow` prints the last N lines first. `databricks air run --watch` still submits and streams. ([#7658](#7658)) * `databricks air logs -o json` and `databricks air run --watch -o json` now emit a `WAITING` event, for example `{"type":"WAITING","message":"Waiting for accelerator compute capacity..."}`, when the reported waiting status changes. ([#7695](#7695)) * `databricks air logs` on a run that ended in the last two minutes now waits for the log index to settle before printing, instead of sometimes dropping lines that were indexed late. ([#7663](#7663)) * `air logs` and `air run --watch` now print the waiting status (for example `Waiting for accelerator compute capacity...`) as a line on stderr when stderr is not a terminal, instead of printing nothing until the run produces logs. ([#7657](#7657)) * Route AIR log links to the MLflow Logs page instead of the hidden artifact path. ([#7491](#7491)) * Invalid `compute.accelerator_type` errors and `air run -h config.compute` no longer list the CLI's built-in accelerator types, which could differ from the workspace's. ([#7520](#7520)) * `databricks air --help` now notes that GPU pools are not required to run workloads. ([#7581](#7581)) * `air run` now prints the job run URL with the `?o=<workspace-id>` suffix, matching `air get` and `air list`; `dashboard_url` in `air run -o json` and the `--watch` events gains it too. When `air list` cannot resolve the workspace ID, its links now omit `?o=` instead of ending in `?o=0`. ([#7573](#7573)) * `databricks air run` now reports an unknown workload YAML key by its field path, suggests the closest valid field, and points to `-h config` instead of naming an internal Go type. ([#7574](#7574)) * `air run` (including `--dry-run`) now warns before submitting when the `code_source` imports `torch` but PyTorch is not declared: the environment is the Standard environment (version 5 or later) with no `torch` dependency, so the run may fail after it waits for a GPU. ([#7656](#7656)) ### Bundles * Add `isolation_mode` to the `catalogs` resource, so a catalog can be deployed as `OPEN` or `ISOLATED`. ([#7463](#7463)) * `bundle validate` no longer warns when resource permissions and top-level or target permissions name the same principal. ([#7291](#7291)) * Retry bundle file uploads, reads and deletes when the server does not respond within the HTTP timeout ("request timed out after ... of inactivity"). ([#7521](#7521)) * Show how to recover when `bundle deploy` fails because an app's service principal does not exist or was deleted. ([#7533](#7533)) * Retry transient server errors when creating Lakebase (`postgres_*`) resources in the direct deployment engine. ([#7557](#7557)) * Fix `bundle plan` reporting the `spark_conf` and `custom_tags` the backend derives from `is_single_node: true` as drift on clusters. ([#7577](#7577)) * Stop planning a spurious update for Lakeflow Connect pipelines whose `ingestion_definition.data_staging_options` is filled in by the backend. ([#7603](#7603)) * Retry the root path check in `bundle destroy` on transient errors, so a dropped connection no longer fails the destroy. ([#7597](#7597)) * Use the Lakebase branch when its create request times out but the branch was created, instead of failing the deploy, in the direct deployment engine. ([#7602](#7602)) * Retry transient server errors when deleting resources in the direct deployment engine. ([#7708](#7708)) * Use the Lakebase project when its create request times out but the project was created, instead of failing the deploy with "already exists", in the direct deployment engine. ([#7747](#7747)) * Use the feature, model serving endpoint or dashboard that a create request created despite timing out, instead of failing the deploy, in the direct deployment engine. ([#7757](#7757)) * Use the Lakebase role when its create request fails with a transient error but the role was created, instead of failing the deploy, in the direct deployment engine. ([#7767](#7767)) * Use the MCP service, model service or model provider service that a create request created despite failing with a timeout, a 499 or a reset stream, instead of failing the deploy, in the direct deployment engine. Create recovery for features, serving endpoints and dashboards now also covers 499 and reset streams. ([#7782](#7782)) * Retry app updates and deletes in the direct deployment engine when they fail with 400 BAD_REQUEST because the app compute is `UPDATING`, after waiting for the compute to settle. ([#7606](#7606)) * Enable drift detection for `source_code_path` and `git_source` on apps. ([#6399](#6399)) * `bundle destroy` and `bundle deploy` no longer fail with "internal error, missing in state" when the state holds an empty-ID entry left by an interrupted recreate in an older CLI. ([#7676](#7676)) * Fix a feature whose `full_name` contains uppercase letters being recreated on every deploy. ([#7398](#7398)) * Fix `bundle deployment bind` (and `bundle generate --bind`) failing with `no such node "resources.internal_immutable_snapshots.immutable"` when `workspace.sync_mode` is `immutable`, as reported in [#7306](#7306). ([#7329](#7329)) * Retry creating the deployment lock (without `--force-lock`) when the request fails with a transient server error, times out ("request timed out after ... of inactivity") or loses its connection ("unexpected EOF"). ([#7594](#7594)) * Fix a materialization set member whose `feature_name` contains uppercase letters failing to deploy, or recreating the set on every deploy. ([#7398](#7398)) * Delete the old resource when its name follows a renamed schema. ([#7718](#7718)) * Fix `pipelines dry-run` starting a real run when given a job KEY, and a crash in its shell completion. ([#7672](#7672)) * Rename schemas in place; schemas and volumes follow a renamed catalog or schema. ([#7343](#7343)) * Fix a panic when validating or deploying jobs with local libraries and serverless job clusters. ([#7725](#7725)) * Fix `tgz` artifacts with `git` set producing an empty tarball or failing when the bundle is in a subdirectory of the git repository. ([#7674](#7674)) * Apply `presets.trigger_pause_status` and the development mode pause default to `triggers[]` entries on jobs. ([#7450](#7450)) ### Dependency Updates * Bump dependencies with known vulnerabilities. ([#7750](#7750)) * Update the Go SDK to [v0.198.0](https://github.com/databricks/databricks-sdk-go/releases/tag/v0.198.0). ([#7530](#7530)) * Update the Go SDK to [v0.199.0](https://github.com/databricks/databricks-sdk-go/releases/tag/v0.199.0). ([#7584](#7584)) * Update the Go SDK to [v0.200.0](https://github.com/databricks/databricks-sdk-go/releases/tag/v0.200.0). ([#7653](#7653)) * Update the Go SDK to [v0.201.0](https://github.com/databricks/databricks-sdk-go/releases/tag/v0.201.0). ([#7769](#7769))
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Title: air: Report each --watch STATUS transition once and never after the run ends
Changes
air run --watch -o jsonemits aSTATUSevent on each lifecycle transition. When the Bricklens stream hands off to the MLflow fallback, the fallback started its own transition tracking from the status resolved before streaming. A run cancelled while PENDING with no logs therefore emitted (current<-previous):With this change all three emit each transition once, in order, ending at the terminal state:
The change has two parts. The tests show that each one is needed.
newStatusTransitionReporter,logstream_support.go; wired atrun.gowhereonStatusChangeis set). It drops a repeat of the last reported state and anything after a terminal state, and it reports the last state it actually emitted asprevious_status. Non-terminal order is not otherwise policed.fetchLogs→mlflowLogFallbackReporting→followMLflowLogs).streamBricklensLogsWithLatestStatusreturns the status the Bricklens loop last observed. The MLflow stream reports that status until its first refresh, so it never surfaces the stale pre-stream one. The fallback's control flow is unchanged: it still follows on the pre-stream status. A run that was being followed keeps the active MLflow stream, with its lenient first pass, 3 s grace wait, re-read, uncapped output, per-chunk offsets and partial-line hold-back. It does not switch to the 10,000-line completed-tail path. The two-pass grace and the three empty terminal Bricklens polls are untouched.The existing
streamBricklensLogs,mlflowLogFallbackandstreamMLflowLogssignatures are kept as thin wrappers, so their callers and tests don't change. Text mode is unaffected; it has no STATUS sink.Why
A report against a workspace:
air run --watchtakes about 60 s to notice a cancel made from another terminal. Investigating it turned up this event-ordering bug. A JSON consumer that treats eachSTATUSas a state machine saw the run go back to PENDING after CANCELED.This patch does not change how quickly a cancel is noticed. A live smoke against a workspace (below) showed the Jobs run state itself staying
RUNNINGfor 67-87 s after a successfulair cancel. Both the patched and the unpatched watch reported CANCELED within about 0.5 s of Jobs flipping, and exited about 2.4 s later. A proposal for the client-side blocking-call case is under "Follow-up design" and isn't implemented here.Tests
cmd/air/run_watch_status_test.go(new). These are full-commandair run --watch -o jsontests againstlibs/testserver. They assert the exactSUBMITTED/STATUSsequence and the closing envelope:TestRunWatchJSONStatusAfterEmptyBricklensHandoff: the run flips PENDING → CANCELED during the first empty workflow-logs poll, and the MLflow IDs appear only on the second fallback lookup. It assertsSUBMITTED, PENDING<-, CANCELED<-PENDING, at least 2 get-output calls (the grace re-read still happens), and all 10,002 lines across two chunks printed (uncapped). The envelope is CANCELED.TestRunWatchJSONStatusAfterFeatureDisabledHandoff: a 403 FEATURE_DISABLED handoff. It assertsSUBMITTED, PENDING<-, CANCELED<-PENDINGand a CANCELED envelope.TestRunWatchJSONStatusAfterActiveMLflowProbeHandoff: MLflow logs are found while RUNNING. It assertsSUBMITTED, PENDING<-, RUNNING<-PENDING, SUCCESS<-RUNNING.TestStatusTransitionReporterDropsRepeatsAndPostTerminalStates: a unit test of the guard.c7d55d3a6, the first two tests fail with the sequences shown above. With only the guard (status not carried), the third test fails with... RUNNING<-PENDING, PENDING<-RUNNING, SUCCESS<-PENDING.go test ./cmd/air/... -count=1(194 s).go test -raceon the touched watch, logstream and MLflow tests passes. A full-race ./cmd/airrun fails only inTestIndexStrategy*(list_index_test.go:43). That failure reproduces on unpatchedc7d55d3a6and isn't touched here.gofmt,golangci-lint run ./cmd/air/...(0 issues) andtools/validate_nextchanges.pypass.Live smoke (against a workspace, 2026-11-13)
One 1×GPU_1xA10
sleep 120workload (max_retries: 0) per build, run withair run --watch -o jsonand cancelled from a second process about 5 s after the first STATUS event:| build | STATUS events | cancel accepted → Jobs TERMINATED/CANCELED | → watch CANCELED event | → exit | exit code |
|---|---|---|---|---|---|---|
| patched |
RUNNING<-,CANCELED<-RUNNING| 87.5 s | +0.2 s | +2.4 s | 1 || unpatched |
RUNNING<-,CANCELED<-RUNNING| 68.2 s | -0.5 s (watch saw it first) | +2.4 s | 1 |Both runs reached RUNNING immediately and printed 7 environment-setup log lines, so this exercised the normal path, not an MLflow handoff. An independent observer polling
jobs get-runandair get -o jsononce per second saw JobsRUNNINGandair getPENDINGfor the whole interval. Both changed to CANCELED at the same moment.Follow-up design (not implemented): an independent status monitor
Under a stalled log request, the watch can't see a cancel. Bricklens is polled serially:
runs/getis refreshed only between iterations, after each workflow-logs GET and its SDK retries. A held request therefore delays noticing the cancel by up to the SDK's 60 s inactivity timeout (a ~68 s delay reproduced locally with an injected stall).Proposal:
We considered and rejected a fixed 8 s deadline on the whole terminal drain, plus an 8 s "status unknown → exit nonzero" rule:
logstream.goandlogmlflow.go.The smoke above suggests the delay users see may be dominated by Jobs cancel propagation, not by the client. It's worth confirming with request traces before investing in the monitor.
This pull request and its description were written by Isaac.