Blueprint structure allows a variety of "illegal" combinations #7078
Assignees
Comments
andrewjstone
pushed a commit
that referenced
this issue
Nov 22, 2024
…ol, it must be from a disk present in the blueprint (#7106) Builds and is staged on top of #7105. The intended change here is in the first commit (8274174): In `BlueprintBuilder::sled_select_zpool()`, instead of only looking at the `PlanningInput`, we also look at the disks present in the blueprint, and only select a zpool that the planning input says is in service and that we have in the blueprint. This had a surprisingly-large blast radius in terms of tests - we had _many_ tests which were adding zones (which implicitly selects a zpool) from a `BlueprintBuilder` where there were no disks configured at all, causing them to emit invalid blueprints. These should all be fixed as of this PR, but I'm a little worried about test fragility in general, particularly with an eye toward larger changes like #7078. Nothing to do about that at the moment, but something to keep an eye on. Fixes #7079.
jgallagher
added a commit
that referenced
this issue
Dec 5, 2024
…uilder` (#7204) This is a step on the road to #7078. The current `BlueprintBuilder` internals would not at all be amenable to squishing the disparate maps in `Blueprint` together, so this PR tries to rework those internals. `BlueprintBuilder` now does its own map squishing (inside `new_based_on()`), combining state+zones+disks+datasets into one helper (`SledEditor`). All modifications are preformed via those editors, then `BlueprintBuilder::build()` breaks the combined editors' results back out into four maps for `Blueprint`. There should be only one functional change in this PR (marked with a `TODO-john`): previously, expunging a zone checked that that zone had not been modified in the current planning cycle. I'm not sure that's really necessary; it felt like some defensiveness born out of the complexity of the builder itself, maybe? I think we could put it back (inside `ZonesEditor`, presumably) if folks think it would be good to keep. My intention was to change the tests as little as possible to ensure I didn't break anything as I was moving functionality around, so `new_based_on()` and `build()` have some arguably-pretty-gross concessions to behave exactly the way the builder did before these changes. I'd like to _remove_ those concessions, but those will be nontrivial behavioral changes that I don't want to try to roll in with all of this cleanup. I think I landed this pretty well; there are only a few expectorate changes (due to the slightly reworked `ZonesEditor` producing a different ordering for a few tests), and none in `omdb` (which is where I've often seen incidental planner changes bubble out). Apologies for the size of the PR. I'd lightly recommend that the new `blueprint_editor/` module and its subcomponents should be pretty quickly reviewable and should be looked at first; they're _supposed_ to be simple and obvious, and are largely ported over from the prior storage / disks / datasets / zones editors. I did rework how we're handling #6645 backwards compat to try to reduce how much we need to pass `SledResources` around, so that complicates things in `DatasetsEditor` some, but not too bad IMO. (I also added a test for this, since I changed it and I don't think we had one before.) The `BlueprintBuilter` changes should then be more or less the natural way one would use a collection of `SledEditor`s without changing its API or behavior (yet!).
|
This shouldn't have been closed; github picked up on "fix #7078" in a commit message, but in context it's clearly not that the commit fixes this. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Blueprintcurrently has four different maps, all keyed by sled ID:omicron/nexus/types/src/deployment.rs
Lines 144 to 163 in 7cf372d
In general we would expect all those maps to have the same keys, but in practice that isn't true. A couple examples:
sled_statedrops decommissioned sleds, but they may still exist in the other maps (but should only contain expunged values). We have to work around this when diff'ing blueprints.BlueprintBuilder, not the planner, that fail to callsled_ensure_disks/sled_ensure_datasetsat appropriate times will emit blueprints that contain zones that reference datasets / disks that one would reasonably expect to exist in their relative maps, but those maps may be out of date or empty. Empty maps happen to work today because for backwards compatibility reasons we allow them to be empty, but doing so is a little messy and will eventually be removed once production systems can all be expected to have populated maps. That will break any of these clients (tests and reconfigurator-cli, today).I think (?) we should probably only have one map here, keyed by sled ID, with values that encompass all the blueprint details for a single sled (today: state + zones + disks + datasets, growing in the future to support update versioning as needed). This may require some fundamental rework of
BlueprintBuilderand possibly the planner, as today they manage these maps mostly independently (which is problematic!).The text was updated successfully, but these errors were encountered: