Oh god I was doing prototyping in this area without bringing in the refactor
Thu, Nov 26
When you say "fixing things", is there a github issue or something that describes what needs to be fixed?
Update: Allow a hover error to be seen and fix all the new URLs
Looks good! It feels like there are some tricky edge cases here, e.g. if only some inputs of a solid come from a prior run, so maybe worth thinking through if there are tests to write that would give us confidence on them.
Wed, Nov 25
just throwing some ideas around -
- we add column to run storage, user_action, user_override or something like that
- add a new event type, not engine event, that corresponds
- run storage response to this new event does a smart column update so that it is only written from "origin host process" ie dagit and shouldn't conflict with other writes
- add a new API (which we could use for other reasons) of instance.get_run_row_by_id to allow fetching the full run row including this column
- use this for cancel / terminate from the dequeuer / run worker
If we do that it should cover terminated too IMO
nit: reported something as "failed" that never started feels weird to me - i think an explicit canceled state may be worth it
oof ya the race conditions here are really gnarly. Probably worth drawing up the full state machine diagram. I feel like we need to have a safer overall scheme like @dgibson is talking about. I think delete and mark-as-failed are emergency escape hatches that may put the system in to a bad state if the user does them at the wrong time. In all other cases there should be a clear "owner" for the run and only that "owner" should manage transitions.
Double check that switching tabs in the editor doesn't create a continuous undo history
This patch package is pretty intriguing. I'm OK with doing this in this isolated instance because we're likely to abandon codemirror at some point in the future (for Monaco) and forking / fixing the library to make this change seems like overkill. I'm curious to get @dish 's thoughts + approval on this though because he investigated that change.
this all seems quite straightforward, we can sort out whether we need an 'Unloadable Sensors' section separately
build completed successfully, not sure why the hook didn't publish.
gql error msg
This seems reasonable to me - curious what @alangenfeld thinks and how this might interact with https://github.com/dagster-io/dagster/issues/3239 and/or https://github.com/dagster-io/dagster/issues/3236 ? Left some thoughts on the race condition inline.
Ok, since we've dropped 2.7, we can go to pickle protocol 4, but not to HIGHEST_PROTOCOL
What's the error message, or does parsing just proceed without the datetime being parsed? YAML is complex enough that I can imagine someone tearing their hair out about this behavior with no clue how to proceed..
My big question is why is this FromMultipleSources and not FromMultipleStepOutputs?
(but also maybe update the summary a bit more?)
I'm not super familiar with this path and it doesn't seem immediately obvious that it is in fact redundant, but it seems plausible and I trust you? Is it because dependencies_dict includes transitive dependencies already? re-request for review if you need more detailed scrutiny