Summary:
Depends on D19078. Ref T13073. Currently, there is a narrow window where we can acquire a resource after a reclaim has started against it.
To prevent this, briefly lock resources before acquiring them and make sure they're still good. If a resource isn't good, throw the lease back in the pool.
Test Plan:
This is tricky. You need:
- Hoax blueprint with limits and a rule where leases of a given "flavor" can only be satisfied by resources of the same flavor.
- Reduce the 3-minute "wait before resources can be released" to 3 seconds.
- Limit Hoaxes to 1.
- Allocate one "cherry" flavored Hoax and release the lease.
- Add a `sleep(15)` to `releaseResource()` in `DrydockResourceUpdateWorker`, after the `canReclaimResource()` check, with a `print`.
Now:
- Run `bin/phd debug task` in two windows.
- Run `bin/drydock lease --type host --attributes flavor=banana` in a third window.
- This will start to reclaim the existing "cherry" resource. Once one of the `phd` windows prints the "RECLAIMING" message run `bin/drydock lease --type host --attributes flavor=cherry` in a fourth window.
- Before patch: the "cherry" lease acquired immediately, then was released and destroyed moments later.
- After patch: the "cherry" lease yields.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13073
Differential Revision: https://secure.phabricator.com/D19080
Summary:
Depends on D19075. Ref T13073. If a lease acquires a resource but finds that the resource builds directly into a dead state (which can happen for any number of reasonable reasons), reset the lease and throw it back in the pool.
This isn't the lease's fault and it hasn't caused any side effects or done anything we can't undo, so we can safely reset it and throw it back in the pool.
Test Plan:
- Created a blueprint which throws from `allocateResource()` so that resources never activate.
- Tried to lease it with `bin/drydock lease ...`.
- Before patch: lease was broken and destroyed after a failed activation.
- After patch: lease was returned to the pool after a failed activation.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13073
Differential Revision: https://secure.phabricator.com/D19076
Summary:
Ref T13073. When a Blueprint says it will be able to allocate a resource but then throws an exception while attempting that allocation, we currently fail the lease permanently.
This is excessively harsh. This blueprint may have the best of intentions and have encountered a legitimately unforseeable failure (like a `vm.new` call to build a VM failed) and be able to succeed in the future.
Even if this blueprint is a dirty liar, other blueprints (or existing resources) may be able to satisfy the lease in the future.
Even if every blueprint is implemented incorrectly, leaving the lease alive lets it converge to success after the blueprints are fixed.
Instead of failing, log the issue and yield.
(In the future, it might make sense to distinguish more narrowly between "actually, all the resources are used up" and all other failure types, since the former is likely more routine and less concerning.)
Test Plan:
- Wrote a broken `Hoax` blueprint which always claims it can allocate but never actually allocates (just `throw` in `allocateResource()`).
- Used `bin/phd drydock lease` to acquire a Hoax lease.
- Before patch: lease abruptly failed permanently.
- After patch: lease yields after allocation fails.
{F5427747}
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13073
Differential Revision: https://secure.phabricator.com/D19070
Summary: Noticed a couple of typos in the docs, and then things got out of hand.
Test Plan:
- Stared at the words until my eyes watered and the letters began to swim on the screen.
- Consulted a dictionary.
Reviewers: #blessed_reviewers, epriestley
Reviewed By: #blessed_reviewers, epriestley
Subscribers: epriestley, yelirekim, PHID-OPKG-gm6ozazyms6q6i22gyam
Differential Revision: https://secure.phabricator.com/D18693
Summary:
Fixes T9994. Currently, when Drydock can't allocate a new resource because some limit has been reached, it waits patiently for a resource to become available.
It is possible that no resource will ever become available. Particularly with "Working Copy" resources, the new lease may want a copy of `rB`, but the resource may already be maxed out on `rA`.
Right now, no process exists to automatically reclaim the unused `rA`.
When we encounter this situation, try to reclaim one of the other resources if it is just sitting there unused.
Specifically:
- Add a "reclaim" command which means "release this resource //if// it is completely unused".
- Add a `bin/drydock reclaim` to send this command to every active resource.
- When we try to acquire a resource and can't, but only because of some kind of limit / utilization problem, try to release an unused resource to free up some room.
Test Plan:
- Set "Working Copy" resource limit to 1.
- Ran "Test Configuration" in `rA`, which worked.
- Ran "Test Configuration" in `rB`, which hung forever.
- Applied patch.
- Ran "Test Configuration" in `rB`, saw it reclaim the `rA` resource, use the slot, then succeed.
- Ran "Test Configuration" in `rA` again, saw it grab the slot back.
- Ran `bin/drydock reclaim` and saw it reclaim a bunch of old orphaned resources.
Reviewers: chad
Reviewed By: chad
Maniphest Tasks: T9994
Differential Revision: https://secure.phabricator.com/D14819
Summary:
Ref T9994. This fixes the first issue discussed on that task, which is that when a merge fails after "arc land", we would not clean up all the leases properly.
Specifically, when a merge fails, we use `queueTask()` to schedule a followup task. This followup destroys the lease and frees the underlying resource.
However, the default behavior of `queueTask()` is to //not queue tasks// if the parent task fails. This is a reasonable, safe behavior that was originally introduced in D8774, where it kept us from sending too much mail if a task did "send some mail" and then failed a little later on and got retried.
Since I think the default behavior is correct, I just special cased the behavior for Drydock to make it queue even on failure. These are the only types of followup tasks we currently want to queue on main task failure.
(It's possible that future Blueprints might want some kind of more specialized behavior, where some tasks queue only on success, but we can cross that bridge when we come to it.)
Test Plan:
- See T9994#149878 for test case setup.
- I ran that test case again with this patch, and saw the followup task queue properly in the `--trace` log, a correspoinding update task show up in `/daemon/`, and the lease get destroyed when I ran it a moment later.
{F1029915}
Reviewers: chad
Reviewed By: chad
Maniphest Tasks: T9994
Differential Revision: https://secure.phabricator.com/D14818
Summary:
Fixes T9669. Two issues:
- We were using `repositoryPHIDs` instead of `blueprintPHIDs` for the list of allowed blueprints. Use the correct value.
- We weren't enforcing `allowedBlueprintPHIDs` fully correctly. We //did// require an authorization, so the net effect was correct in nearly all cases, but we could have selected from too large a pool in the case where the application itself was doing the authorization (e.g., from the command line).
Test Plan: Ran a build through Drydock/Harbormaster locally.
Reviewers: chad, tycho.tatitscheff
Reviewed By: chad, tycho.tatitscheff
Subscribers: tycho.tatitscheff
Maniphest Tasks: T9669
Differential Revision: https://secure.phabricator.com/D14368
Summary:
Ref T182. Couple of minor improvements here:
- Show the Drydock lease when viewing a Repository Operation detail screen. This just makes it easier to jump around between relevant objects.
- When tasks are waiting for a lease, awaken them when it breaks or is released, not just when it is acquired. This makes the queue move forward faster when errors occur.
Test Plan:
- Viewed a repository operation and saw a link to the lease.
- Did a bad land (intentional merge problem) and got an error in about ~3 seconds instead of ~17.
Reviewers: chad
Reviewed By: chad
Maniphest Tasks: T182
Differential Revision: https://secure.phabricator.com/D14341
Summary:
Ref T9252. This is mostly a fix for an edge case from D14236. Here's the setup:
- There are no resources.
- A request for a new resource arrives.
- We build a new resource.
Now, if we were leasing an existing resource, we'd call `canAcquireLeaseOnResource()` before acquiring a lease on the new resource.
However, for new resources we don't do that: we just acquire a lease immediately. This is wrong, because we now allow and expect some resources to be unleasable when created.
In a more complex workflow, this can also produce the wrong result and leave the lease acquired sub-optimally (and, today, deadlocked).
Make the "can we acquire?" pathway consistent for new and existing resources, so we always do the same set of checks.
Test Plan:
- Started daemons.
- Deleted all working copy resources.
- Ran two working-copy-using build plans at the same time.
- Before this change, one would often [1] acquire a lease on a pending resource which never allocated, then deadlock.
- After this change, the same thing happens except that the lease remains pending and the work completes.
[1] Although the race this implies is allowed (resource pool limits are soft/advisory, and it is expected that we may occasionally run over them), it's MUCH easier to hit right now than I would expect it to be, so I think there's probably at least one more small bug here somewhere. I'll see if I can root it out after this change.
Reviewers: chad, hach-que
Reviewed By: hach-que
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14272
Summary:
Ref T9519. When acquiring leases on resources:
- Only consider resources created by authorized blueprints.
- Only consider authorized blueprints when creating new resources.
- Fail with a tailored error if no blueprints are allowed.
- Fail with a tailored error if missing authorizations are causing acquisition failure.
One somewhat-substantial issue with this is that it's pretty hard to figure out from the Harbormaster side. Specifically, the Build step UI does not show field value anywhere, so the presence of unapproved blueprints is not communicated. This is much more clear in Drydock. I'll plan to address this in future changes to Harbormaster, since there are other related/similar issues anyway.
Test Plan: {F872527}
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T9519
Differential Revision: https://secure.phabricator.com/D14254
Summary:
Ref T9252. I think there's a more complex version of this problem discussed elsewhere, but here's what we hit today:
- 5 commits land at the same time and trigger 5 builds.
- All of them go to acquire a working copy.
- Working copies have a limit of 1 right now, so 1 of them gets the lease on it.
- The other 4 all trigger allocation of //new// working copies. So now we have: 1 active, leased working copy and 4 pending, leased working copies.
- The 4 pending working copies will never activate without manual intervention, so these 4 builds are stuck forever.
To fix this, prevent WorkingCopies from giving out leases until they activate. So now the leases won't acquire until we know the working copy is good, which solves the first problem.
However, this creates a secondary problem:
- As above, all 5 go to acquire a working copy.
- One gets it.
- The other 4 trigger allocations, but no longer acquire leases. This is an improvement.
- Every time the leases update, they trigger another allocation, but never acquire. They trigger, say, a few thousand allocations.
- Eventually the first build finishes up and the second lease acquires the working copy. After some time, all of the builds finish.
- However, they generated an unboundedly large number of pending working copy resources during this time.
This is technically "okay-ish", in that it did work correctly, it just generated a gigantic mess as a side effect.
To solve this, at least for now, provide a mechanism to impose allocation rate limits and put a cap on the number of allocating resources of a given type. As hard-coded, this the greater of "1" or "25% of the active resources in the pool".
So if there are 40 working copies active, we'll start allocating up to 10 more and then cut new allocations off until those allocations get sorted out. This prevents us from getting runaway queues of limitless size.
This also imposes a total active working copy resource limit of 1, which incidentally also fixes the problem, although I expect to raise this soon.
These mechanisms will need refinement, but the basic idea is:
- Resources which aren't sure if they can actually activate should wait until they do activate before allowing leases to acquire them. I'm fairly confident this rule is a reasonable one.
- Then we limit how many bookkeeping side effects Drydock can generate once it starts encountering limits.
Broadly, some amount of mess is inevitable because Drydock is allowed to try things that might not work. In an extreme case we could prevent this mess by setting all these limits at "1" forever, which would degrade Drydock to effectively be a synchronous, blocking queue.
The idea here is to put some amount of slack in the system (more than zero, but less than infinity) so we get the performance benefits of having a parallel, asyncronous system without a finite, manageable amount of mess.
Numbers larger than 0 but less than infinity are pretty tricky, but I think rules like "X% of active resources" seem fairly reasonable, at least for resources like working copies.
Test Plan:
Ran something like this:
```
for i in `seq 1 5`; do sh -c '(./bin/harbormaster build --plan 10 rX... &) &'; done;
```
Saw 5 plans launch, acquire leases, proceed in an orderly fashion, and eventually finish successfully.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14236
Summary:
Ref T9252. This primarily allows Harbormaster to request (and Drydock to fulfill) working copies with a patch from a staging area. Doing this means we can do builds on in-review changes from `arc diff`.
This is a little cobbled-together but should basically work.
Also fix some other issues:
- Yielded, awakend workers are fine to update but could complain.
- We can't log slot lock failures to resources if we don't end up saving them.
- Killing the transaction would wipe out the log.
- Fix some TODOs, etc.
Test Plan: Ran Harbormaster builds on a local revision.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14214
Summary:
Ref T9252. Long ago you sometimes manually created resources, so they had human-enterable names. However, users never make resources manually any more, so this field isn't really useful any more.
In particular, it means we write a lot of untranslatable strings like "Working Copy" to the database in the default locale. Instead, do the call at runtime so resource names are translatable.
Also clean up a few minor things I hit while kicking the tires here.
It's possible we might eventually want to introduce a human-choosable label so you can rename your favorite resources and this would just be a default name. I don't really have much of a use case for that yet, though, and I'm not sure there will ever be one.
Test Plan:
- Restarted a Harbormaster build, got a clean build.
- Released all leases/resources, restarted build, got a clean build with proper resource names.
Reviewers: hach-que, chad
Reviewed By: hach-que, chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14213
Summary:
Ref T9252. See companion change in D14211. This does the same thing for leases.
Particularly, most of the TODOs about error handling can just be removed because they'll do the right things by default now.
This and D14211 also move slot lock release to after resource destruction. This feels cleaner than trying to release early at release/break.
Test Plan: Restarted a Harbormaster build, got a clean build result. This needs more vetting but I'll clean up any issues as I hit them.
Reviewers: chad, hach-que
Reviewed By: hach-que
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14212
Summary:
Ref T9252. This is the same as D14201, but for lease stuff instead of resource stuff.
This one is a little heavier but still feels pretty reasonable to me at the end of the day (worker is <1K lines and has a ton of comment stuff).
Also fixes a few random bugs I hit in the task queue.
Test Plan:
- Restarted some Harbormaster builds, saw them go through cleanly.
- Released pre-activation resources/leases.
- Probably still kinda buggy but I'll iron the details out over time.
Logs are starting to look somewhat plausible:
{F855747}
Reviewers: chad
Reviewed By: chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14202
Summary: Ref T9252. Make log types modular so they can be translated and have complicated rendering logic if necessary (currently, none have this).
Test Plan: {F855330}
Reviewers: chad
Reviewed By: chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14198
Summary:
Fixes T6569. This implements an expiry mechanism for Drydock resources which parallels the mechanism for leases.
A few things are missing that we'll probably need in the future:
- An "EXPIRES" command to update the expiration time. This would let resources be permanent while leased, then expire after, say, 24 hours without any leases.
- A callback like `shouldActuallyExpireRightNow()` for resources and leases that lets them decide not to expire at the last second.
- A callback like `didAcquireLease()` for resource blueprints, to parallel `didReleaseLease()`, letting them clear or extend their timer.
However, this stuff would mostly just let us tune behaviors, not really open up new capabilities.
Test Plan: Changed host resources to expire after 60 seconds, leased one, saw it vanish 60 seconds later.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T6569
Differential Revision: https://secure.phabricator.com/D14176
Summary: Ref T6569. If a lease is activated with an expiration date, schedule a task to try to clean it up after that time.
Test Plan:
- Used `bin/drydock lease ... --until ...` to activate a lease in the near future.
- Waited for a bit.
- Saw it expire and get destroyed at the scheduled time.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T6569
Differential Revision: https://secure.phabricator.com/D14148
Summary: Ref T9252. Some leases or resources may need to remove data, tear down VMs, etc., during cleanup. After they are released, queue a "destroy" phase for performing teardown.
Test Plan:
- Used `bin/drydock lease ...` to create a working copy lease.
- Used `bin/drydock release-lease` and `bin/drydock release-resource` to release the lease and then the working copy and host.
- Saw working copy and host get destroyed and cleaned up properly.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T6569, T9252
Differential Revision: https://secure.phabricator.com/D14144
Summary:
Ref T9252. Broadly, Drydock currently races on releasing objects from the "active" state. To reproduce this:
- Scatter some sleep()s pretty much anywhere in the release code.
- Release several times from web UI or CLI in quick succession.
Resources or leases will execute some release code twice or otherwise do inconsistent things.
(I didn't chase down a detailed reproduction scenario for this since inspection of the code makes it clear that there are no meaningful locks or mechanisms preventing this.)
Instead, add a Harbormaster-style command queue to resources and leases. When something wants to do a release, it adds a command to the queue and schedules a worker. The workers acquire a lock, then try to consume commands from the queue.
This guarantees that only one process is responsible for writes to active resource/leases.
This is the last major step to giving resources and leases a single writer during all states:
- Resource, Unsaved: AllocatorWorker
- Resource, Pending: ResourceWorker (Possible rename to "Allocated?")
- Resource, Open: This diff, ResourceUpdateWorker. (Likely rename to "Active").
- Resource, Closed/Broken: Future destruction worker. (Likely rename to "Released" / "Broken"; maybe remove "Broken").
- Resource, Destroyed: No writes.
- Lease, Unsaved: Whatever wants the lease.
- Lease, Pending: AllocatorWorker
- Lease, Acquired: LeaseWorker
- Lease, Active: This diff, LeaseUpdateWorker.
- Lease, Released/Broken: Future destruction worker (Maybe remove "Broken"?)
- Lease, Expired: No writes. (Likely rename to "Destroyed").
In most phases, we can already guarantee that there is a single writer without doing any extra work. This is more complicated in the "Active" case because the release buttons on the web UI, the release tools on the CLI, the lease requestor itself, the garbage collector, and any other release process cleaning up related objects may try to effect a release. All of these could race one another (and, in many cases, race other processes from other phases because all of these get to act immediately) as this code is currently written. Using a queue here lets us make sure there's only a single writer in this phase.
One thing which is notable is that whatever acquires a lease **can not write to it**! It is never the writer once it queues the lease for activation. It can not write to any resources, either. And, likewise, Blueprints can not write to resources while acquiring or releasing leases.
We may need to provide a mechinism so that blueprints and/or resource/lease holders get to attach some storage to resources/leases for bookkeeping. For example, a blueprint might need to keep some kind of cache on a resource to help it manage state. But I think we can cross that bridge when we come to it, and nothing else would need to write to this storage so it's technically straightforward to introduce such a mechanism if we need one.
Test Plan:
- Viewed buttons in web UI, checked enabled/disabled states.
- Clicked the buttons.
- Saw commands show up in the command queue.
- Saw some daemon stuff get scheduled.
- Ran CLI tools, saw commands get consumed and resources/leases release.
Reviewers: hach-que, chad
Reviewed By: chad
Maniphest Tasks: T9252
Differential Revision: https://secure.phabricator.com/D14143
