| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
| |
Also clarifies a comment on preventNewEvaluations.
RELNOTES: None.
PiperOrigin-RevId: 184198568
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In the case that a node is already done when evaluation starts, we now report
events and postables early, rather than waiting until the end of evaluation.
This makes reporting more timely, and ensures reporting even if the evaluation
is interrupted.
This caused a problem with moving the TargetCompleteEvent into Skyframe
(unknown commit). I added two unit tests at the Skyframe level to cover the
guarantees that we need for that.
Note that the replay call in constructResult can duplicate the events from
a cache hit - this is not a problem since the replaying visitor automatically
removes duplicates (and it wasn't obvious which keys correspond to cache
hits).
PiperOrigin-RevId: 179788157
|
|
|
|
|
|
|
|
| |
Blaze had its own class to avoid GC from varargs array creation for the precondition happy path. Guava now (mostly) implements these, making it unnecessary to maintain our own.
This change was almost entirely automated by search-and-replace. A few BUILD files needed fixing up since I removed an export of preconditions from lib:util, which was all done by add_deps. There was one incorrect usage of Preconditions that was caught by error prone (which checks Guava's version of Preconditions) that I had to change manually.
PiperOrigin-RevId: 175033526
|
|
|
|
|
|
| |
This change prepares a move to Guava's preconditions. Guava only has vararg-avoiding overloads up to 4 args.
PiperOrigin-RevId: 175015502
|
|
|
|
| |
PiperOrigin-RevId: 174202685
|
|
|
|
|
|
|
|
| |
catastrophes. Our stricter behavior in the face of errors means that it is no longer possible for a done node to depend on a not-done node in this build.
This opens up the possibility to discard graph edges on all --batch builds, or at least those with --discard_analysis_cache.
PiperOrigin-RevId: 171375405
|
|
|
|
|
|
|
|
|
|
| |
This change enables alternate evaluation strategies.
Drive-by fix to GraphTester, making TestFunction static because it has
no dependencies on its outer class, and adding #unsetConstantValue.
RELNOTES: None.
PiperOrigin-RevId: 164327055
|
|
|
|
|
|
|
| |
A subsequent CL makes TargetPatternKey implement SkyKey, and it's much nicer
if I can pass lists of TargetPatternKey to the various Skyframe APIs.
PiperOrigin-RevId: 162359843
|
|
|
|
|
|
|
| |
We were previously duplicate-posting Postable events posted to the
Skyframe environment.
PiperOrigin-RevId: 162323598
|
|
|
|
|
|
|
|
|
|
| |
-Have SkylarkImportLookupFunction include causes in the SkyFunctionExceptions it throws.
-Better transitive skyframe error declarations in ASTFileLookupFunction.
-Have ErrorInfo differentiate between direct and transitive transience.
-Introduce ErrorInfoManager and have ParallelEvaluator/ParallelEvaluatorContext use it.
RELNOTES: None
PiperOrigin-RevId: 159163186
|
|
|
|
|
|
|
|
|
|
|
|
| |
With the introduction of the ExtendedEventHandler, SkyFunctions were
given the possibility to post additional Postable events in addition
to the standard events. As SkyFunctions have to be restartable, events
are collected first and only posted after the function is finished.
Make sure that this also applies to postable events and they are not
dropped.
Change-Id: Ie1c3a0134935c75ea984fa2cc924e7327a9da81f
PiperOrigin-RevId: 158964337
|
|
|
|
|
| |
RELNOTES: None.
PiperOrigin-RevId: 156553687
|
|
|
|
|
| |
RELNOTES: None
PiperOrigin-RevId: 155665128
|
|
|
|
|
|
| |
when change pruning. This can speed up change pruning.
PiperOrigin-RevId: 152538144
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
reverseDepsToConsolidate field in InMemoryNodeEntry. As part of that, revamp our logic of how we store pending operations: store adds bare on initial evaluations, and checks bare on incremental evaluations and operations on done nodes.
This should improve performance in two ways: BuildingState loses two fields, saving working memory intra-build. Storing pending reverse dep operations bare also saves memory intra-build. Note that neither of these changes helps resting memory state, only while a node is still evaluating.
Because of this, we can simplify ReverseDepsUtil a bit, making ReverseDepsUtilImpl a static class, which it always wanted to be (what it really wants to be is a superclass of InMemoryNodeEntry, but I don't want to spend the object alignment bits).
Finally, this makes it pretty tempting to get rid of BuildingState altogether on initial evaluations. We'd still keep DirtyBuildingState, but we could save another ~24 bytes by storing BuildingState's one remaining field, signaledDeps, directly inside InMemoryNodeEntry.
--
PiperOrigin-RevId: 151048879
MOS_MIGRATED_REVID=151048879
|
|
|
|
|
|
|
|
| |
a child that's being checked by a parent (that was a legacy of when we delegated to enqueueChild), and such a child that is being checked should always be dirty or done, never fresh.
--
PiperOrigin-RevId: 149136909
MOS_MIGRATED_REVID=149136909
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
With more specific information to be reported by Skyfunctions, e.g.,
to inform the build-event protocol on missing files, the EventHandler
interface is no longer enough. Therefore, provide an enriched context
for reporting events.
--
Change-Id: I2d06166fe4d5b9054e24ad8c752fafc039e3f9f8
Reviewed-on: https://cr.bazel.build/8794
PiperOrigin-RevId: 148463437
MOS_MIGRATED_REVID=148463437
|
|
|
|
|
|
|
|
| |
handling.
--
PiperOrigin-RevId: 147185674
MOS_MIGRATED_REVID=147185674
|
|
|
|
|
|
|
|
| |
useful, but may also help identify Bazel as the culprit in the linked bug if we're lucky and the timestamps are clear.
--
PiperOrigin-RevId: 144359628
MOS_MIGRATED_REVID=144359628
|
|
|
|
|
|
|
| |
Instead, deduplicate when the helper is actually added to a GroupedList.
--
MOS_MIGRATED_REVID=136145321
|
|
|
|
|
|
|
|
|
|
| |
Remove DirtyKeyTracker and inflight node tracking as independent
progress tracking bits of code, and instead use an internal-facing
DirtyTrackingProgressReceiver wrapper of the user's
EvaluationProgressReceiver.
--
MOS_MIGRATED_REVID=135291924
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=132999234
|
|
|
|
|
|
|
| |
can use different cycle detection algorithms if they wish.
--
MOS_MIGRATED_REVID=131347160
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=131344134
|
|
|
|
|
|
|
|
|
|
|
| |
respect to cycle checking.
Reducing the size of ParallelEvaluator.java is also probably long overdue.
I believe this change stands on its own, but if you don't think the third change is worth it, and this isn't worth it on its own, feel free to push back.
--
MOS_MIGRATED_REVID=131340165
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Instead, just put them directly into a map. This avoids the memory churn and CPU cost of the set.
As a result, we have to use HashMaps instead of ImmutableMap.Builders, which I hope is ok (especially since we're not keeping them around), and due to that, we have some nice nondeterminism in the returned order, which matters for some cycle-checking tests.
Also, don't use a map at all when we don't need to (when building events).
Note that, since we have to deduplicate at some point, this means that changing the return type of SkyFunction.Environment#getValues to not be a random-access map is probably not worth it. Changing the return type of ProcessableGraph#getBatch to not be a random access map might still be worthwhile, although it might require some funny operations.
--
MOS_MIGRATED_REVID=131070418
|
|
|
|
|
|
|
| |
The only place we now don't handle InterruptedException is in the action graph created after analysis, since I'm not sure that will be around for that much longer.
--
MOS_MIGRATED_REVID=130327770
|
|
|
|
|
|
|
|
|
| |
Make sure that ParallelEvaluator treats SchedulerExceptions as less severe than other RuntimeExceptions (it already did, but add a comment emphasizing this).
The combination of the above means that we don't ignore hypothetical crashes in ParallelEvaluator worker threads _after_ a SchedulerException (e.g. due to an error in nokeep_going evaluation).
--
MOS_MIGRATED_REVID=130044230
|
|
|
|
|
|
|
|
|
| |
nokeep_going build because it finished building in between the time it was first requested and when we checked it for done-ness after the SkyFunction evaluation.
This results in an IllegalStateException that gets ignored (and, importantly, not propagated) by AbstractQueueVisitor because AbstractQueueVisitor only records and propagates the first Throwable encountered. For nokeep_going evaluations, this will be the SchedulerException that we use for control flow. The IllegalStateException in question is benign in this case because it's merely from a Preconditions failure and doesn't leave anything in a bad state. It's possible, though, that we have other bugs that are being masked in this way.
--
MOS_MIGRATED_REVID=129919336
|
|
|
|
|
|
|
| |
QueryableGraph.Reason which conveys more information. Add a few more Reason enum values to make this refactor benign.
--
MOS_MIGRATED_REVID=129118462
|
|
|
|
|
|
|
| |
parameters conveying the requesting node (if any), the requested node(s), as well as a reason for the skyframe graph lookup. Alternate graph implementations may be interested in this information.
--
MOS_MIGRATED_REVID=128496089
|
|
|
|
|
|
|
| |
SkyFunction if it throws an exception but still has missing deps. Instead, pretend it didn't throw, and restart it when its known deps are all done, presumably to throw the same exception. This removes a basic source of non-determinism.
--
MOS_MIGRATED_REVID=127129202
|
|
|
|
|
|
|
| |
alternate graph implementations to optimize how they construct node entries.
--
MOS_MIGRATED_REVID=126932020
|
|
|
|
|
|
|
| |
last build's deps. This is only an issue with cycles, since there we try to maintain the invariant that a parent is not done before its children by removing its deps on its children before we construct its cycle value.
--
MOS_MIGRATED_REVID=126494009
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=126344189
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=125376554
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=125362963
|
|
|
|
|
|
|
|
|
|
|
| |
nodes, where there is no work to do anyway.
This triggered some non-determinism that we explicitly workaround in the unit tests.
Also add a comment about a potential but unrelated optimization.
--
MOS_MIGRATED_REVID=125355303
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=124874002
|
|
|
|
|
|
|
| |
functionality was only used in tests.
--
MOS_MIGRATED_REVID=124841573
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=123347295
|
|
|
|
|
|
|
| |
where we weren't checking to see if a reverse dep already existed when we declared a reverse dep.
--
MOS_MIGRATED_REVID=122581019
|
|
|
|
|
|
|
|
|
|
|
| |
discover that it no longer has certain deps.
In the common case, where a node's deps do not change in the end, this reduces lock contention and CPU.
The downside of this is that we now create a set of the previous reverse deps during each evaluation of a node. We don't store this set in order to conserve memory, so we pay for it in CPU. We will probably only construct it two or three times (most SkyFunctions don't have so many groups), so the cost shouldn't be so high, but we can try to mitigate if it shows up in profiling.
--
MOS_MIGRATED_REVID=122566267
|
|
|
|
|
|
|
| |
conversion is unnecessary and wasteful. In the remaining cases, the set conversion can be explicit.
--
MOS_MIGRATED_REVID=122294939
|
|
|
|
|
|
|
| |
helpers to enforce concurrency synchronization points and determinism even if they are not using an InMemoryGraph-backed evaluator.
--
MOS_MIGRATED_REVID=121977783
|
|
|
|
|
|
|
| |
non-keep-going build. Setting errorDepKey was only mostly harmless, to mix memes. (Actually, it was quite harmful.)
--
MOS_MIGRATED_REVID=118410594
|
|
|
|
|
| |
--
MOS_MIGRATED_REVID=117894693
|
|
|
|
|
|
|
| |
it. The copying showed up as a source of memory spikiness.
--
MOS_MIGRATED_REVID=117741939
|
|
|
|
|
|
|
| |
are created, as opposed to when they are requested from the ParallelEvaluator. That delay can lead to large memory spikes and churn.
--
MOS_MIGRATED_REVID=116224565
|
|
|
|
|
|
|
| |
SkyFunction
--
MOS_MIGRATED_REVID=115671161
|