- Bottom line - we need to manage shutting down gracefully when doing long-running tasks
- To achieve that, we're going to use job queues
In this commit:
- added new @tryghost/job-manager dependency
- added a minimal job service, that handles in passing things like logging and (maybe later) config
- job service is wired up to server shutdown, so that the queue finishes before the server exits
- also added a new job endpoint to testmode so that it's easy to test job behaviour without needing to do real work
- Using consistent patterns for shutdown and stop
- Make it clear what each method does
- Use async/await to make the code more readable and simple
- This lays groundwork for having more cleanup tasks in stop than just server.stop()
- deleted files under `core/server/lib/promise` and related test files
- added `@tryghost/promise` as a dependency
- fixed all local requires to point to the new package
closes#12118
- server.start was mistakenly removed in 71f02d25e9
- it is used for loading themes (and other things) and is critical
- added tests to prevent this regressing again in future
no issue
- New Member API batched import is meant to be a substitution to current import
with improved performance while keeping same behaviore. Current
import processes 1 record at a time using internal API calls and times
out consistently when large number of members has to be imported (~10k
records without Stripe).
- New import's aim is to improve performance and process >50K
records without timing out both with and without Stripe connected
members
- Batched import can be conceptually devided into 3 stages which have
their own ways to improve performance:
1. labels - can be at current performance as number of
labels is usually small, but could also be improved through batching
2. member records + member<->labels relations - these could
be performed as batched inserts into the database
3. Stripe connections - most challanging bottleneck to solve because
API request are slow by it's nature and have to deal with rate limits of
Stripe's API itself
- It's a heavy WIP, with lots of known pitfalls which are marked with
TODOs. Will be solved iteratively through time untill the method can be
declared stable
- The new batched import method will be hidden behind 'enableDeveloperExperiments' flag to
allow early testing
- A simple way to test behaviours without having to do complex interactions to e.g. generate errors or slow requests
- Makes it easier to test the new shutdown behaviour, among other things
- there can now be quite a big delay between SIGINT/TERM being received and shutdown finishing
- add an extra log message to acknowledge the SIGINT/TERM to facilitate debugging and just be clear
- stopppable is a dependency that handles closing connections properly, which server.close does not
- active connections are allowed to complete what they are doing
- idle connections are closed
- no new connections are allowed
- we call stoppable in stop() instead of server.close so that idle connections don't hold the server open
- calling await stop() from shutdown then ensures that we have a consistent experience of stop
- all together this allows ghost to shutdown gracefully when there are long-running requests
- @TODO: handle graceful shutdown of long-running processes
- @TODO: consider do we need to send 503s whilst the server is shutting down?
- changed method to logStopMessages, as we use start and stop, not start and shutdown
- changed logStopMesasges to output the "proper" messages and use this method consistently - the closing connections message isn't really useful
- changed uptime message to always be output cos I can't see a case where there isn't interesting/useful
- Connection handling is legacy code added in 1438278ce4
- Although we were tracking all connections in memory, we weren't actually closing any because stop isn't called
- This (and restart) were both added as part of the now long-deprecated system for using Ghost directly as an npm module
- If we want to close connections cleanly, we should use a tool to do this
