Ensure all errors in `cli` inherit from `ApplicationError` to continue
normalizing all the errors we report to Sentry
Follow-up to: https://github.com/n8n-io/n8n/pull/7820
Followup to #7566 | Story: https://linear.app/n8n/issue/PAY-926
### Manual workflow activation and deactivation
In a multi-main scenario, if the user manually activates or deactivates
a workflow, the process (whether leader or follower) that handles the
PATCH request and updates its internal state should send a message into
the command channel, so that all other main processes update their
internal state accordingly:
- Add to `ActiveWorkflows` if activating
- Remove from `ActiveWorkflows` if deactivating
- Remove and re-add to `ActiveWorkflows` if the update did not change
activation status.
After updating their internal state, if activating or deactivating, the
recipient main processes should push a message to all connected
frontends so that these can update their stores and so reflect the value
in the UI.
### Workflow activation errors
On failure to activate a workflow, the main instance should record the
error in Redis - main instances should always pull activation errors
from Redis in a multi-main scenario.
### Leadership change
On leadership change...
- The old leader should stop pruning and the new leader should start
pruning.
- The old leader should remove trigger- and poller-based workflows and
the new leader should add them.
- Enable two-way communication with web sockets
- Enable sending push messages to specific users
- Add collaboration service for managing active users for workflow
Missing things:
- State is currently kept only in memory, making this not work in
multi-master setups
- Removing a user from active users in situations where they go inactive
or we miss the "workflow closed" message
- I think a timer based solution for this would cover most edge cases.
I.e. have FE ping every X minutes, BE removes the user unless they have
received a ping in Y minutes, where Y > X
- FE changes to be added later by @MiloradFilipovic
Github issue / Community forum post (link here to close automatically):
---------
Co-authored-by: कारतोफ्फेलस्क्रिप्ट™ <aditya@netroy.in>
https://linear.app/n8n/issue/PAY-933/set-up-leader-selection-for-multiple-main-instances
- [x] Set up new envs
- [x] Add config and license checks
- [x] Implement `MultiMainInstancePublisher`
- [x] Expand `RedisServicePubSubPublisher` to support
`MultiMainInstancePublisher`
- [x] Init `MultiMainInstancePublisher` on startup and destroy on
shutdown
- [x] Add to sandbox plans
- [x] Test manually
Note: This is only for setup - coordinating in reaction to leadership
changes will come in later PRs.
This change ensures that things like `encryptionKey` and `instanceId`
are always available directly where they are needed, instead of passing
them around throughout the code.
This PR adds a message for queue mode which triggers an external secrets
provider reload inside the workers if the configuration has changed on
the main instance.
It also refactors some of the message handler code to remove cyclic
dependencies, as well as remove unnecessary duplicate redis clients
inside services (thanks to no more cyclic deps)
all commands sent between main instance and workers need to contain a
server id to prevent senders from reacting to their own messages,
causing loops
this PR makes sure all sent messages contain a sender id by default as
part of constructing a sending redis client.
---------
Co-authored-by: कारतोफ्फेलस्क्रिप्ट™ <aditya@netroy.in>
This PR implements the updated license SDK so that worker and webhook
instances do not auto-renew licenses any more.
Instead, they receive a `reloadLicense` command via the Redis client
that will fetch the updated license after it was saved on the main
instance
This also contains some refactoring with moving redis sub and pub
clients into the event bus directly, to prevent cyclic dependency
issues.
# Motivation
In Queue mode, finished executions would cause the main instance to
always pull all execution data from the database, unflatten it and then
use it to send out event log events and telemetry events, as well as
required returns to Respond to Webhook nodes etc.
This could cause OOM errors when the data was large, since it had to be
fully unpacked and transformed on the main instance’s side, using up a
lot of memory (and time).
This PR attempts to limit this behaviour to only happen in those
required cases where the data has to be forwarded to some waiting
webhook, for example.
# Changes
Execution data is only required in cases, where the active execution has
a `postExecutePromise` attached to it. These usually forward the data to
some other endpoint (e.g. a listening webhook connection).
By adding a helper `getPostExecutePromiseCount()`, we can decide that in
cases where there is nothing listening at all, there is no reason to
pull the data on the main instance.
Previously, there would always be postExecutePromises because the
telemetry events were called. Now, these have been moved into the
workers, which have been given the various InternalHooks calls to their
hook function arrays, so they themselves issue these telemetry and event
calls.
This results in all event log messages to now be logged on the worker’s
event log, as well as the worker’s eventbus being the one to send out
the events to destinations. The main event log does…pretty much nothing.
We are not logging executions on the main event log any more, because
this would require all events to be replicated 1:1 from the workers to
the main instance(s) (this IS possible and implemented, see the worker’s
`replicateToRedisEventLogFunction` - but it is not enabled to reduce the
amount of traffic over redis).
Partial events in the main log could confuse the recovery process and
would result in, ironically, the recovery corrupting the execution data
by considering them crashed.
# Refactor
I have also used the opportunity to reduce duplicate code and move some
of the hook functionality into
`packages/cli/src/executionLifecycleHooks/shared/sharedHookFunctions.ts`
in preparation for a future full refactor of the hooks
This PR adds new endpoints to the REST API:
`/orchestration/worker/status` and `/orchestration/worker/id`
Currently these just trigger the return of status / ids from the workers
via the redis back channel, this still needs to be handled and passed
through to the frontend.
It also adds the eventbus to each worker, and triggers a reload of those
eventbus instances when the configuration changes on the main instances.
Issue: during startup, unfinished executions trigger a recovery process
that, under certain circumstances, can in itself crash the instance
(e.g. by running our of memory), resulting in an infinite recovery loop
This PR aims to change this behaviour by writing a flag file when the
recovery process starts, and removing it when it finishes. In the case
of a crash, this flag will persist and upon the next attempt, the
recovery will instead do the absolute minimal (marking executions as
'crashed'), without attempting any 'crashable' actions.
