A shard server is impractical without operation subscription, partition
subscription confirmation between servers and workers.
The failsafe execution is configuration that is likely not desired on
workers. This change removes the failsafe behavior from workers via
backplane config, and relegates the setting of failsafe boolean to
server config. If the option is restored for workers, it can be added to
worker configs so that configs may continue to be shared between workers
and servers and retain independent addressability.

Internally driven metrics and scaling controls have low, if any, usage
rates. Prometheus has largely succeeded independent publication of
metrics, and externally driven scaling is the norm. These modules have
been incomplete between cloud providers, and for the functional side of
AWS, bind us to springboot. Removing them for the sake of reduced
dependencies and complexity.

Remove this unused OperationQueue feature which provides no invocations
on any use.

This reverts commit f651cdb.

Continue the loop while we have *not* matched successfully and avoid a
confusing inversion in getMatched()

Distinguish the valid/unique/propagating methods of entry listening.

The only signal to a waiting match that will halt its current listen
loop for a valid unique operation is an interrupt.

Distinguish target param with GRPC type storage from FILESYSTEM
definition

Reinstate prior usage of LoggingMain for safe shutdown, with added
release mechanism for interrupted processes. All invoked shutdowns are
graceful, with vastly improved shutdown speed for empty workers waiting
for pipeline stages.

Tiny code cleanup

Will include operation root and inform directory cache effectiveness.

Selecting realInputDirectories by regex permits flexible patterns that
can yield drastic improvements in directory reuse for specialized
deployments. runfiles in particular are hazardous expansions of
nearly-execroot in the case of bazel.

Care must be taken to match directories exclusively.
The entire input tree is traversed for matches against expanded paths
under the root, to allow for nested selection.
Each match thus costs the number of input directories.
Counterintuitively, OutputFiles are augmented to avoid the recursive
check for OutputDirectories which only applies to actual reported
results, resulting in a path match when creating the exec root.
Regex style is java.util.Pattern, and must match the full input
directory.

This will include the path to the missed directory and the operation
which required it.

Prevent adding duplicate realInputDirectories matches

Ensure that the last leg of the execution presents a directory, rather
than the parent, per OutputDirectory's stamping.

Support a `--redis_uri` command line option for start-up.

also fixed some spelling typos.

Bumping 4.2 -> 5.3

Bump fro 0.0.6 -> 0.0.9

Don't get transitive grpc dependencies, use the ones from our `maven_install(...)`

To aid connection troubleshooting

buildfarm#1486)

* Add unit test
* Signal Write on complete

Closes buildfarm#1508

Cleanups:
- remove the unused `ubuntu-bionic` base image
- replace `ubuntu-jammy:jammy-java11-gcc` with `ubuntu-mantic:mantic-java17-gcc`
- replace `amazoncorretto:19` with `ubuntu-mantic:mantic-java17-gcc`
- swap inverted log file names in a file

Specifying any number of OutputPaths will ignore OutputFiles (consistent
with uploads). Where an OutputPath specifies an output directory, the
action must be able to create the directory itself.

Partial specification of the absolute symlink response per REAPI.
Remaining work will be in output identification.

Trying to get more info on the Lombok stamping issue on Windows CI.
See also bazelbuild/bazel#10363 and
bazelbuild/bazel#18185

Default disabled, available with createSymlinkOutputs option in Worker
config.

Problem

    Enabling the findMissingBlobsViaBackplane flag in BuildfarmServer eliminates the need for the BuildfarmWorker's fmb API call. This BuildfarmWorker:fmb call was also responsible for tracking CAS entry access. As result, our CAS cache eviction strategy shifted from LRU to FIFO.
    When the findMissingBlobsViaBackplane flag is enabled, the buildfarm relies on the backplane as the definitive source for CAS availability. Since we don't update CAS expiry on each access, the backplane will independently expire CAS entries based on the specified cas_expire duration, even if they are actively being read.

Solution

Updated bfServer:fmb call to perform non-blocking fmb calls to workers, allowing these workers to record access for the relevant CAS entries.

Extended expiry duration for available CAS entries in the backplane on each fmb call.

With these changes, we can utilize Bazel's experimental_remote_cache_lease_extension and experimental_remote_cache_ttl flags for incremental builds.

Closes buildfarm#1428

…m#1516)

Bumps [org.json:json](https://github.com/douglascrockford/JSON-java) from 20230227 to 20231013.
- [Release notes](https://github.com/douglascrockford/JSON-java/releases)
- [Changelog](https://github.com/stleary/JSON-java/blob/master/docs/RELEASES.md)
- [Commits](https://github.com/douglascrockford/JSON-java/commits)

---
updated-dependencies:
- dependency-name: org.json:json
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <[email protected]>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

Specifying a correlated invocation id with a uri containing a
THROUGH=true query param will cause the CFC to read a blob through an
external input stream, populating locally along the way. This permits
client-based replication of blobs, and can enable N+1 replication and
traffic balancing for reads.

Option to run the worker with a cmdline specification for its gRPC
server port.

Root cmdline specification has been broken since the config change of
v2.

Remove the size identification for small blobs when uploading with
bf-executor.

Permit installations to control the failure process for operations which
produce outputs larger than the maxEntrySizeBytes. A default value
of false retains the existing behavior which appears transient and
blacklists the executed action key. When enabled, the action will fail
under an invalid violation that indicates user error.

There's been a few releases of it by now and this pulls the latest. For
buildfarm, notable changes included performance enhancments during
decompression.

See:
https://github.com/facebook/zstd/releases/tag/v1.5.5

Update for further docs related to storage+type functionality
Remove outdated Operation Queue worker definitions

When a write fails because the write was already cancelled before due to something like deadline exceeded, we get an unknown error. The exception comes from here and when it gets to errorResponse(), it only checks if status code is cancelled. In this case the status code is unknown, so we need to check if context is cancelled to prevent responseObserver from being invoked

The code change adds checking if context is cancelled and a unit test testing when the exception has context cancelled.

Release notes: https://github.com/google/error-prone/releases/tag/v2.22.0

tmpfs
Previously we required bazel clients to opt-in with an exec_property. Now buildfarm has the ability to enforce tmpfs usage as part of deployment configuration.
See: alwaysUseTmpFs: true

cgroups
cgroups was previously enabled via limitGlobalExecution. However, there was no way to opt-out of cgroups and opt-in to other restrictions. limitGlobalExecution is still required for accessing restrictions, but it will no longer guarantee cgroups is used. For that, deployments must set:
See: alwaysUseCgroups: true

Additional context
We may also want to experiment with sandbox, but not cgroups in our deployment. There is currently no way to do this because enabling sandbox also requires enabling limitGlobalExecution, but doing so enables cgroups. By making these settings more granular we are able opt into and enforce particular execution wrappers.

Refer to new camelized DMS fields.
Expand predefined dynamic execution property name matches.

…uildfarm#1530)

Add an option to provide a list of custom label headers to add to metrics.

Testing with bazel HEAD using jdk21 compilation has revealed new
direct dependencies on guava.

Annotations under lombok were fixed for jdk21 in 1.18.28, update to
current.

Remove acceptEverything DequeueMatchSetting
Place worker name in workerProvisions
Only enable allowUnmatched effects on key mismatch
Only acquire resources after asserting compatibility
Update documentation to match changes

Correct deprecated AccessController usage warning
Requires a newer bazel than 6.4.0 for macos to choose unix toolchain with C++ std=c++14 specification for protobuf->absl dependency.

Create a github workflow to build base buildfarm worker image.

The execution platform property "cores" is detailed in documentation as
specifying "min-cores" and "max-cores". Match this definition and
prevent "cores" from being evaluated as a strict match with the worker
provision properties (with likely rejection).

Per the REAPI spec:

`The paths are relative to the working directory of the action
execution.`

Prefix the WorkingDirectory to paths used as OutputDirectory parameters,
and verify that these are present in the layout of the directory for
use.

Followup to buildfarm#1195

Add a new execution pathway in worker/Executor.java to use persistent workers via PersistentExecutor, like DockerExecutor.

Mostly unchanged from the form we used to experiment back at Twitter, but now with tests.

Co-authored-by: Shane Delmore [email protected]

* Locate Output Paths relative to WorkingDirectory

Required as a corollary to OutputDirectory changes to consider outputs
as relative to working directory.

* Windows builds emit relativize paths with native separators

Previous patch included a change in actionRoot parameter, expecting it
to prefer the working directory rooted path to discover outputs. Might
want to reapply this later, but for now leave the resolution in
uploadOutputs.

* Prevent multiple fetches of QueuedOperation

Propagate the Tree in OperationContext for use during the persistent
workers clause in Executor

* Correct WSO and StubWriteOutputStream completion

onCompleted should always be called on open requests when
StubWriteOutputStreams are closed, even when completed early by the
server.
onCompleted must be delivered to responseObserver when onCompleted is
invoked, lest a no-request write stream be left hanging to
DEADLINE_EXCEEDED.

* Explicit reset on WriteOffset 0 and blob write

Avoid unnecessary queryWriteStatus requests when a WriteRequest
appears with write offset 0 in WSO, and reset the stream initially for
internal ServerInstance blob uploads.

… artifact (buildfarm#1556)

* bundle with helm
publish release with gh cli

* add a basic doc for the helm chart