Odara 0.4.1
Released 2026-11-17. A fix release for the Monitor's record of Maestro runs.
Fixed
- A Maestro run started from the API (
POST /maestros/:id/run) or by a
schedule recorded no steps in the Monitor. Every way of starting a Maestro
now records its steps the same way.
- A Maestro run from the editor was recorded with 0/0 pipelines. The
completed, failed and total counts are now recorded for every run.
- The child pipelines a Maestro ran locally had no per-node rows in the
Monitor (isolated children did). A local run now records each node, like a
run started from the editor.
- The Monitor did not link a Maestro run to its child runs. Each pipeline step
now names the run it started and each child run names its Maestro run and
step; the run the editor announces is the one recorded. In the Monitor, a
Maestro run lists its steps, and "Open run →" shows a child run's nodes.
Tip
A parallel group that calls the same pipeline twice in a Local environment
runs both calls on the same server. If the pipeline writes a fixed path,
the two calls write the same file: give the path a variable the call sets,
e.g. /tmp/out_@{PART}.csv.
Odara 0.4.0
Released 2026-11-16. Linux (tar.gz, .deb, .rpm) and Windows (zip), installed with https://odara.rs/install.sh and install.ps1.
New
- Isolated runs. Each environment chooses where its runs happen:
- Local, inside the Odara server, as before. This is the default.
- Docker: one container per run on the server, with CPU, memory and time limits.
- AWS ECS Fargate: one task per run in your AWS account.
- Google Cloud Run: one job execution per run in your Google Cloud project.
The machine is deleted at the end of the run. AWS and Google support is in every download. Setup is in deploy/aws/ (a CloudFormation template) and deploy/gcp/ (setup.sh); the guide is docs/isolated-runs/USER-GUIDE.md.
- Secret values never travel in the run bundle. On AWS and Google, each run gets its own vault entry, deleted when the run ends.
- The Odara server only gets the rights to start, watch and stop its runs and to read their logs.
- A pipeline that needs the server's disk is refused before launch, with its paths listed. A run may use its own
/tmp for a file that one node writes or downloads and another reads or uploads. - The Monitor records an isolated run like a local one: per-node rows and times, and the target history. The editor shows each node running, then done.
- Abort and the time limit stop the run. Out of memory is reported as such.
- Choose the environment per run. The editor's top bar lists the environments, with a badge saying where each one runs; the choice belongs to the pipeline's tab and applies to Execute and Preview.
- The API takes
environment on /pipelines/:id/run, /run-stream and /maestros/:id/execute. - A Maestro run's child pipelines run in its environment.
- The Monitor shows each run's environment.
- The server's active environment no longer has to change.
- Schedules have an environment. Left empty, a schedule uses the active environment when it fires, as before.
- Manage environments in Admin › Environments: make one active, create, rename (the id used by variables never changes), delete (its variables go with it). Base is not listed there; it holds the values shared by all environments.
- Google nodes in isolated runs. A Google node's
credentials_path may hold the service-account key JSON, normally through a secret variable (@{GCP_KEY}), instead of a key file on the server. - New logo in the app header (larger), the sign-in page and the browser tab.
- New default theme, Marble: white with cool greys and the violet accent. It replaces Atlântico (a saved Atlântico choice becomes Marble). The theme names are now in English: Marble, Porcelain, Dark.
Behavior changes (read before upgrading)
- File sources and targets refuse remote URLs. A CSV, Excel, XML or Magic File path such as
s3://…, https://… or sftp://… was never read; it now fails at once, naming the node to use (S3 Download/Upload, Azure Blob, FTP Get/Put). Before, an s3:// target created a local folder named s3:. - BigQuery Target writes with a load job, instead of streaming inserts. Load jobs are free and see a table recreated by
drop_and_create at once. - Linear pipelines that take the streaming fast path (for example PostgreSQL → SQL → PostgreSQL) now record target history and per-node events, like other runs.
Fixed
- Rerunning a pipeline that recreates its BigQuery table failed with "table not found" for minutes.
- A Python step killed by the memory limit failed with an empty message. It now says it most likely ran out of memory.
- In the light themes, the options of native dropdowns (Create Schedule and others) were white on white.
- Switching an environment's Execution to another mode and back gave the defaults instead of the values entered.
- Deleting an environment left its variables behind, and other environments could still promote to it.
- There was no way in the UI to change the active environment.
- Google Drive Copy in batch (source file from a column) failed: the run used an old single-file copy.
- An isolated run whose client disconnected left its Monitor record "running"; the run now finishes and is recorded on its own. A Docker run whose log stream hangs after the container exits no longer waits forever.
- The editor's edges kept their "data flowing" glow after a run ended when another tab's pipeline was running; they now follow only the open tab.
- The canvas hand cursor almost disappeared on white backgrounds; it is now a dark hand with a white halo.
- Some white texts on light backgrounds in the light themes (Create Schedule type buttons, the AI chat title), and primary buttons that stayed blue.
Validation
Checked against real services on 2026-11-16 (docs/isolated-runs/RESULTS.md):
| What | Result |
|---|
| Docker isolated runs through the server (tests/isolated-runs/e2e_product.py) | 13/13 |
| AWS ECS Fargate (e2e_aws.py): S3 → CSV → SQL → RDS, Monitor, editor progress, abort, time limit, out of memory, nothing left | 9/9 |
| Every S3 node on Fargate (e2e_aws_s3_nodes.py): List, Download, Upload, Copy, Delete, single and batch | 10/10 |
| Google Cloud Run (e2e_gcp.py): cross-cloud S3 → SQL → S3, abort, time limit, out of memory, nothing left | 8/8 |
| BigQuery Source/Target, Local and on Cloud Run (e2e_gcp_nodes.py) | 6/6 |
| Every Google Drive node, single and batch, Local and on Cloud Run, user OAuth (e2e_drive.py) | 16/16 |
| Concurrent isolated runs, 2/4/8 at once, including clients that hang up (stress_concurrent.py) | 40/40 rounds |
| Same pipeline at once in two environments (local and container), and schedules with and without an environment | 11/11 |
| Same pipeline at once in DEV (Local), STG (Google Cloud Run) and PROD (AWS Fargate), via the per-run environment | 3/3 |
| No secret in any response, Monitor, CloudWatch or Cloud Logging entry checked | yes |
Automated tests: odarax-api 257 passed (product build, with AWS and Google); odarax-scheduler 14 passed; odarax-core 369 passed, plus 1 known failure of the test machine (a_python_replacement_is_tested_and_built, no pyarrow).
Google Drive from a service account: a service account cannot create files in a personal Drive (no storage quota). Use the user's OAuth, as tested, or a Shared Drive.