HotLoop Cascade 1.4.0 release notes
What this is
Section titled “What this is”1.4.0 makes a Cascade install easier to run for a long time: the server waits for its database instead of crashing, finished runs can be kept for a number of days and cleaned up with their files, an owner can cap how often one API key calls, and the whole instance can be backed up and put back. It also settles the one scale test the roadmap had left open. The engine, the workflow file format, the REST API, the MCP server (83 tools) and the 14 languages are what 1.3.0 had.
Waits for its database
Section titled “Waits for its database”In the split install (server, worker and webhook pods on PostgreSQL), the pods used to start before the database’s name resolved and restarted two or three times before they were ready. 1.2.1 and 1.3.0 did the same. A server, worker or webhook that can’t reach its database now says so in its log (“waiting for the database”, with how long it has waited) and tries again, for up to five minutes, then connects when the database answers. A wrong password, a malformed address or a database that doesn’t exist still stops it at once, because waiting can’t fix those. The chart’s startup probe waits as long as the program does. A fresh split install on a real k3s showed no restarts, in all three of the chart test’s runs.
Keeping finished runs
Section titled “Keeping finished runs”Settings has a new field, Keep finished runs for (days), and PATCH /api/v1/settings takes retentionDays. It is off by default (0), and every run is kept as it always was. Set a number and once an hour a main deletes the succeeded, failed and canceled runs that ended longer ago than that, with their step data. A run that is still going is never touched, nor is a run whose parent is. Each batch writes execution.pruned to the audit log, and changing the setting writes instance.retention.
The same hourly pass now cleans up the blob store, whether retention is on or not: output too big for the database, and attachments, that no run points at any more. A file two runs share stays until the last of them goes, and only files written more than 24 hours ago are looked at, so a file a step has just stored is never removed. This is new for runs you delete by hand too; before, a deleted run left its big outputs on disk for good. A pass that can’t read what it needs removes nothing, and each pass writes blobs.collected to the audit log.
A limit for each API key
Section titled “A limit for each API key”Settings has a new field, API requests per key per minute, and PATCH /api/v1/settings takes apiRateLimit. It is off by default. Set a number and each API key, and each agent signed in over MCP with one, may make that many requests a minute. A call over the limit is answered 429 with a Retry-After. Each key has its own count, and people signed in to the editor are not held to it. The count is each process’s own, so with three mains a key can make up to three times the number. Changing it writes instance.api_rate_limit to the audit log.
Backup and restore
Section titled “Backup and restore”hotloop-cascade backup DIR writes a new directory with a consistent copy of the SQLite database taken while the server runs, the key file that seals every stored credential, the blob files and a manifest.json of their hashes. hotloop-cascade restore DIR puts them back, after checking every file against the manifest, and refuses to overwrite a database or a different key unless you say --replace. For PostgreSQL the backup holds the key, the blobs and the manifest, and prints the pg_dump command for the database’s half; restore prints the pg_restore one. Blobs kept in S3 or Azure are not copied. docs/GOVERNANCE.md has the details.
The tests that stood behind it
Section titled “The tests that stood behind it”- The scale test the roadmap set is done. Two mains on one PostgreSQL ran a workflow that fires every minute for an hour while one of them was killed with SIGKILL at the slot boundary, seven times in turn and left down about 25 seconds each time: 61 runs for 61 minute slots, none missed, none doubled.
- The test stack’s OAuth server (Hydra) now runs on Postgres. It failed in parallel runs before, passed after, and failed again with the change reverted. No Cascade code introspects tokens.
- The chart test now fails if a pod restarts at install.
- A log destination that fails now waits out its back-off from when the failure is on record, not from when the send failed. A slow write of the failure note used up the first second, and the retry straight after was sent instead of held back; the suite stopped on it under load.
- The language check now sweeps the screen that reads one run, at both widths, in every language, because a run’s step and status words are the ones a plant reads most.
- The two new settings refusals are translated in every language, and the settings help says the big outputs go with their runs.
amd64 only, and why
Section titled “amd64 only, and why”This release builds amd64 only. Multi-arch ships in its own run, so the release after this one is a single small version bump that builds both. stable, latest, 1.4 and 1 list the new amd64 image beside the arm64 image they already pointed at.
Upgrading, and going back
Section titled “Upgrading, and going back”No migrations: the schema is the one 1.3.0 has. helm upgrade to 1.4.0, or let :stable roll the Quadlet units. Putting 1.3.1 back is safe, because nothing in the database changed. What a rollback can’t undo is a prune: runs that retention deleted are gone, and so are the blobs the hourly cleanup removed. Retention is off until you turn it on, so take a backup first; the blob cleanup runs without it.