* Fix a bug speedUpSimulation is not enabled
If the connectionFailureDisableTime is slightly larger than connectionFailureDisableTime,
disableConnectionFailures returns a small value such that disableConnectionFailuresAfter()
is not retrying. As a result, clogging is still in place, and cause many test failures.
A typical symptom is that some transaction can't commit due to transaction_too_old errors.
If we look closely, we'll find that there are many ProxyReject errors with QDelay > 5s.
A failure I looked at has only one commit proxy and one resolver, and the latency
between them is quite high:
5.494145 Sim2Connection Machine=2.0.2.1:2 ID=0000000000000000 From=2.0.2.1:2 To=2.0.2.2:2 SendBufSize=1736851 Latency=0.0292578 StableConnection=0
5.494145 Sim2Connection Machine=2.0.2.1:2 ID=0000000000000000 From=2.0.2.2:2 To=2.0.2.1:2 SendBufSize=510278 Latency=0.0158367 StableConnection=0
So after a while, all transactions got either transaction_too_old errors or
ProxyReject'ed, because resolution took about 100ms on average.
To reproduce:
Commit: a3720567a2
Seed: -b on -f tests/fast/BackupCorrectnessWithEKPKeyFetchFailures.toml -s 2458961283
* Add a const DISABLE_CONNECTION_FAILURE_MIN_INTERVAL
* Fix isOnMainThread in Simulation and Testing
isOnMainThread() is used to check if the currently running task
is on the FDB's event loop. However, in simulation this behaviour
is broken and always returns false.
In other modes such as UnitTest mode since `runTests()` is called before
`g_network->run()`, but without a wait() statement the event loop never
gets chance to set itself as main thread and the tests never sees
current thread as main thread. Therefore we add a yield inside
`runTests()` so yield control back to caller block and continue
with g_network->run() which eventually schedule it back after
initialization.
* Update fdbserver/tester.actor.cpp
Co-authored-by: Syed Paymaan Raza <1238752+spraza@users.noreply.github.com>
---------
Co-authored-by: Syed Paymaan Raza <1238752+spraza@users.noreply.github.com>
The Tester may disable the connection failure after the GcGenerations enables
it. So we want to extend the connection failure for the Tester in this case.
* Ensuring HTTP ports are stable for the same servers, and adding a test http server that ensures no other users accidentally talk to it
* fixing warning with werror
* more werror fixes
* [fdbserver] workaround the FRT type layout issue to get Swfit getVersion working
* MasterData.actor.h: fix comment typo
* masterserver.swift: some tweaks
* masterserver.swift: remove getVersion function, use the method
* masterserver.swift: print replied version to output for tracing
* [swift] add radar links for C++ interop issues found in getVersion bringup
* Update fdbserver.actor.cpp
* Migrate MasterData closer to full reference type
This removes the workaround for the FRT type layout issue, and gets us closer to making MasterData a full reference type
* [interop] require a new toolchain (>= Oct 19th) to build
* [Swift] fix computation of toAdd for getVersion Swift implementation
* add Swift to FDBClient and add async `atLeast` to NotifiedVersion
* fix
* use new atLeast API in master server
* =build fixup link dependencies in swift fdbclient
* clocks
* +clock implement Clock using Flow's notion of time
* [interop] workaround the immortal retain/release issue
* [swift] add script to get latest centos toolchain
* always install swift hooks; not only in "test" mode
* simulator - first thing running WIP
* cleanups
* more cleanup
* working snapshot
* remove sim debug printlns
* added convenience for whenAtLeast
* try Alex's workaround
* annotate nonnull
* cleanup clock a little bit
* fix missing impls after rebase
* Undo the swift_lookup_Map_UID_CommitProxyVersionReplies workaround
No longer needed - the issue was retain/release
* [flow][swift] add Swift version of BUGGIFY
* [swiftication] add CounterValue type to provide value semantics for Counter types on the Swift side
* remove extraneous requestingProxyUID local
* masterserver: initial Swift state prototype
* [interop] make the Swiftied getVersion work
* masterserver - remove the C++ implementation (it can't be supported as state is now missing)
* Remove unnecessary SWIFT_CXX_REF_IMMORTAL annotations from Flow types
* Remove C++ implementation of CommitProxyVersionReplies - it's in Swift now
* [swift interop] remove more SWIFT_CXX_REF_IMMORTAL
* [swift interop] add SWIFT_CXX_IMMORTAL_SINGLETON_TYPE annotation for semanticly meaningful immortal uses
* rename SWIFT_CXX_REF_IMMORTAL -> UNSAFE_SWIFT_CXX_IMMORTAL_REF
* Move master server waitForPrev to swift
* =build fix linking swift in all modules
* =build single link option
* =cmake avoid manual math, just get "last" element from list
* implement Streams support (#18)
* [interop] update to new toolchain #6
* [interop] remove C++ vtable linking workarounds
* [interop] make MasterData proper reference counted SWIFT_CXX_REF_MASTERDATA
* [interop] use Swift array to pass UIDs to registerLastCommitProxyVersionReplies
* [interop] expose MasterServer actor to C++ without wrapper struct
* [interop] we no longer need expose on methods 🥳
* [interop] initial prototype of storing CheckedContinuation on the C++ side
* Example of invoking a synchronous swift function from a C++ unit test. (#21)
* move all "tests" we have in Swift, and priority support into real modules (#24)
* Make set continuation functions inline
* Split flow_swift into flow_swift and flow_swift_future to break circular dependency
* rename SwiftContinuationCallbackStruct to FlowCallbackForSwiftContinuation
* Future interop: use a method in a class template for continuation set call
* Revert "Merge pull request #22 from FoundationDB/cpp-continuation" (#30)
* Basic Swift Guide (#29)
Co-authored-by: Alex Lorenz <arphaman@gmail.com>
* Revert "Revert "Merge pull request #22 from FoundationDB/cpp-continuation" (#30)"
This reverts commit c025fe6258.
* Restore the C++ continuation, but it seems waitValue is broken for CInt somehow now
* disable broken tests - waitValue not accessible
* Streams can be async iterated over (#27)
Co-authored-by: Alex Lorenz <arphaman@gmail.com>
* remove work in progress things (#35)
* remove some not used (yet) code
* remove expose func for CInt, it's a primitive so we always have witness info (#37)
* +masterdata implement provideVersions in Swift (#36)
* serveLiveCommittedVersion in Swift (#38)
* Port updateLiveCommittedVersion to swift (#33)
Co-authored-by: Konrad `ktoso` Malawski <konrad_malawski@apple.com>
* Implement updateRecoveryData in Swift (#39)
Co-authored-by: Alex Lorenz <arphaman@gmail.com>
* Simplify flow_swift to avoid multiple targets and generate separate CheckedContinuation header
* Uncomment test which was blocked on extensions not being picked up (#31)
* [interop] Use a separate target for Swift-to-C++ header generation
* reduce boilerplate in future and stream support (#41)
* [interop] require interop v8 - that will fix linker issue (https://github.com/apple/swift/issues/62448)
* [interop] fix swift_stream_support.h Swift include
* [interop] bump up requirement to version 9
* [interop] Generalize the Flow.Optional -> Swift.Optional conversion using generics
* [WIP] masterServer func in Swift (#45)
* [interop] Try conforms_to with a SWIFT_CONFORMS_TO macro for Optional conformance (#49)
* [interop] include FlowOptionalProtocol source file when generating Flow_CheckedContinuation.h
This header generation step depends on the import of the C++ Flow module, which requires the presence of FlowOptionalProtocol
* conform Future to FlowFutureOps
* some notes
* move to value() so we can use discardable result for Flow.Void
* make calling into Swift async funcs nicer by returning Flow Futures
* [interop] hide initial use of FlowCheckedContinuation in flow.h to break dependency cycle
* [fdbserver] fix an EncryptionOpsUtils.h modularization issue (showed up with modularized libc++)
* Pass GCC toolchain using CMAKE_Swift_COMPILE_EXTERNAL_TOOLCHAIN to Swift's clang importer
* [interop] drop the no longer needed libstdc++ include directories
* [cmake] add a configuration check to ensure Swift can import C++ standard library
* [swift] include msgpack from msgpack_DIR
* [interop] make sure the FDB module maps have 'export' directive
* add import 'flow_swift' to swift_fdbserver_cxx_swift_value_conformance.swift
This is needed for CONFORMS_TO to work in imported modules
* make sure the Swift -> C++ manually bridged function signature matches generated signature
* [interop][workaround] force back use of @expose attribute before _Concurrency issue is fixed
* [interop] make getResolutionBalancer return a pointer to allow Swift to use it
We should revert back to a reference once compiler allows references again
* [interop] add a workaround for 'pop' being marked as unsafe in Swift
* masterserver.swift: MasterData returns the Swift actor pointer in an unsafe manner
* Add a 'getCopy' method to AsyncVar to make it more Swift friendly
* [interop] bump up the toolchain requirement
* Revert "[interop][workaround] force back use of @expose attribute before _Concurrency issue is fixed"
This reverts commit b01b271a76.
* [interop] add FIXME comments highlighting new issue workarounds
* [interop] adopt the new C++ interoperability compiler flag
* [interop] generate swift compile commands
* Do not deduplicate Swift compilation commands
* [interop] generate swift compile commands
* Do not deduplicate Swift compilation commands
* flow actorcompiler.h: add a SWIFT_ACTOR empty macro definition
This is needed to make the actor files parsable by clangd
* [cmake] add missing dependencies
* experimental cross compile
* [cmake] fix triple in cross-compiled cmake flags
* [interop] update to interop toolchain version 16
* [x-compile] add flags for cross-compiling boost
* cleanup x-compile cmake changes
* [cmake] fix typo in CMAKE_Swift_COMPILER_EXTERNAL_TOOLCHAIN config variable
* [interop] pass MasterDataActor from Swift to C++ and back to Swift
* [fdbserver] Swift->C++ header generation for FDBServer should use same module cache path
* Update swift_get_latest_toolchain.sh to fetch 5.9 toochains
* set HAVE_FLAG_SEARCH_PATHS_FIRST for cross compilation
* Resolve conflicts in net2/sim2/actors, can't build yet
* undo SWIFT_ACTOR changes, not necessary for merge
* guard c++ compiler flags with is_cxx_compile
* Update flow/actorcompiler/ActorParser.cs
Co-authored-by: Evan Wilde <etceterawilde@gmail.com>
* update the boost dependency
* Include boost directory from the container for Swift
* conform flow's Optional to FlowOptionalProtocol again
* Guard entire RocksDBLogForwarder.h with SSD_ROCKSDB_EXPERIMENTAL to avoid failing on missing rocksdb APIs
* remove extraneous merge marker
* [swift] update swift_test_streams.swifto to use vars in more places
* Add header guard to flow/include/flow/ThreadSafeQueue.h to fix moduralization issue
* Update net and sim impls
* [cmake] use prebuilt libc++ boost only when we're actually using libc++
* [fdbserver] Swift->C++ header generation for FDBServer should use same module cache path
* fixups after merge
* remove CustomStringConvertible conformance that would not be used
* remove self-caused deprecation warnings in future_support
* handle newly added task priority
* reformatting
* future: make value() not mutating
* remove FIXME, not needed anymore
* future: clarify why as functions
* Support TraceEvent in Swift
* Enable TraceEvent using a class wrapper in Swift
* prearing WITH_SWIFT flag
* wip disabled failing Go stuff
* cleanup WITH_SWIFT_FLAG and reenable Go
* wip disabled failing Go stuff
* move setting flag before printing it
* Add SWIFT_IDE_SETUP and cleanup guides and build a bit
* Revert "Wipe packet buffers that held serialized WipedString (#10018)"
This reverts commit e2df6e3302.
* [Swift] Compile workaround in KeyBackedRangeMap; default init is incorrect
* [interop] do not add FlowFutureOps conformance when building flow clang module for Flow checked continuation header pre-generation
* make sure to show -DUSE_LIBCXX=OFF in readme
* readme updates
* do not print to stderr
* Update Swift and C++ code to build with latest Swift 5.9 toolchain now that we no longer support universal references and bridge the methods that take in a constant reference template parameter correctly
* Fix SERVER_KNOBS and enable use them for masterserver
* Bump to C++20, Swift is now able to handle it as well
* Put waitForPrev behind FLOW_WITH_SWIFT knob
* Forward declare updateLiveCommittedVersion
* Remove unused code
* fix wrong condition set for updateLiveCommittedVersion
* Revert "Revert "Wipe packet buffers that held serialized WipedString (#10018)""
This reverts commit 5ad8dce052.
* Enable go-bindings in cmake
* Revert "Revert "Wipe packet buffers that held serialized WipedString (#10018)""
This reverts commit 5ad8dce052.
* USE_SWIFT flag so we "build without swift" until ready to by default
* uncomment a few tests which were disabled during USE_SWIFT enablement
* the option is WITH_SWIFT, not USE
* formatting
* Fix masterserver compile error
* Fix some build errors.
How did it not merge cleanly? :/
* remove initializer list from constructor
* Expect Swift toolchain only if WITH_SWIFT is enabled
* Don't require Flow_CheckedContinuation when Swift is disabled
* Don't compile FlowCheckedContinuation when WITH_SWIFT=OFF
* No-op Swift macros
* More compile guards
* fix typo
* Run clang-format
* Guard swift/bridging include in fdbrpc
* Remove printf to pass the test
* Remove some more printf to avoid potential issues
TODO: Need to be TraceEvents instead
* Remove __has_feature(nullability) as its only used in Swift
* Don't use __FILENAME__
* Don't call generate_module_map outside WITH_SWIFT
* Add some more cmake stuff under WITH_SWIFT guard
* Some more guards
* Bring back TLSTest.cpp
* clang-format
* fix comment formatting
* Remove unused command line arg
* fix cmake formatting in some files
* Address some review comments
* fix clang-format error
---------
Co-authored-by: Alex Lorenz <arphaman@gmail.com>
Co-authored-by: Russell Sears <russell_sears@apple.com>
Co-authored-by: Evan Wilde <etceterawilde@gmail.com>
Co-authored-by: Alex Lorenz <aleksei_lorenz@apple.com>
Co-authored-by: Vishesh Yadav <vishesh_yadav@apple.com>
Co-authored-by: Vishesh Yadav <vishesh3y@gmail.com>
* Passes existing tests
* adding http unit test for wrong md5 sum
* Added new HTTPKeyValueStore workload to test long-running http clients
* fixing warnings
* Enable secure allocation mode in Arena
This mode allows zeroing out blocks holding sensitive data after use
* Introduce WipedString to all token-holding memory
Also introduce a option flag "sensitive"
* Make pointer equivalency a hard requirement for non-ASAN builds
So that we can detect when Arena/malloc/memory-wipe behavior changes
When remote DC is down, the remote team collection of DD can initializing
waiting for the remote to recover (all_tlog_recruited state). However, the
getTeam request can already be served by the remote team collection. So, for
a RelocateShard (data movement such as split, move), it will get a team for
the remote DC. But the data movement can't make progress on the remote team
because the remote DC hasn't recovered yet. Because of the stuck of data
movement, the primary cannot reach the "storage_recovered" state and stay in
accepting_commit state.
The specifc test failure: slow/ApiCorrectness.toml -s 339026305 -b on
at commit: 0edd899d65
In this test, primary DC has 1 SS killed, remote DC has 2 TLog and 2 SS killed.
So the remote is dead, the remaining 2 SSes can't make progress because of the
loss of 2 TLogs. The repairDeadDatacenter() can't reach the "storage_recovered"
state due to DD's failure of moving shards away from the killed SS in the
primary.
The fix is to exclude all remote in repairDeadDatacenter() so that tells DD to
mark all SSes in the remote as unhealthy. Another fix is to return empty
results for getTeam request if the remote team collection is not ready. This
will allow the data movement to continue, essentially remote team is not changed
for the data movement.
Also, to minimize audit log loss, handle token usage audit logging at each usage.
This has a side-effect of making the token use log less bursty.
This also subtly changes the dedup cache policy.
Dedup time window used to be 5 seconds (default) since the start of batch-logging.
Now it's 5 seconds from the first usage since the closing of the previous dedup window
In the HA configuration, it's possible the remote DC was killed 2 out of 3
machines, left not enough machines for a successful recovery. So this PR changes
to Reboot to avoid such excessive killings.
- Added a background actor that listens on METRICS_EMISSION_UDP_PORT for incoming metrics (and verifies they are in the correct format)
- TraceEvent details have certain requirements for naming. This commit makes a seperate name for Counter/LatencySample and its underlying IMetric to avoid
those issues
Jingyu Zhou