This is a major change which handles 400 error responses when uploading
large files from memory heap/buffer (or staging committer) and the remote S3
store returns a 500 response from a upload of a block in a multipart upload.
The SDK's own streaming code seems unable to fully replay the upload;
at attempts to but then blocks and the S3 store returns a 400 response
"Your socket connection to the server was not read from or written to
within the timeout period. Idle connections will be closed.
(Service: S3, Status Code: 400...)"
There is an option to control whether or not the S3A client itself
attempts to retry on a 50x error other than 503 throttling events
(which are independently processed as before)
Option: fs.s3a.retry.http.5xx.errors
Default: true
500 errors are very rare from standard AWS S3, which has a five nines
SLA. It may be more common against S3 Express which has lower
guarantees.
Third party stores have unknown guarantees, and the exception may
indicate a bad server configuration. Consider setting
fs.s3a.retry.http.5xx.errors to false when working with
such stores.
Signification Code changes:
There is now a custom set of implementations of
software.amazon.awssdk.http.ContentStreamProvidercontent in
the class org.apache.hadoop.fs.s3a.impl.UploadContentProviders.
These:
* Restart on failures
* Do not copy buffers/byte buffers into new private byte arrays,
so avoid exacerbating memory problems..
There new IOStatistics for specific http error codes -these are collected
even when all recovery is performed within the SDK.
S3ABlockOutputStream has major changes, including handling of
Thread.interrupt() on the main thread, which now triggers and briefly
awaits cancellation of any ongoing uploads.
If the writing thread is interrupted in close(), it is mapped to
an InterruptedIOException. Applications like Hive and Spark must
catch these after cancelling a worker thread.
Contributed by Steve Loughran
1. The class WrappedIO has been extended with more filesystem operations
- openFile()
- PathCapabilities
- StreamCapabilities
- ByteBufferPositionedReadable
All these static methods raise UncheckedIOExceptions rather than
checked ones.
2. The adjacent class org.apache.hadoop.io.wrappedio.WrappedStatistics
provides similar access to IOStatistics/IOStatisticsContext classes
and operations.
Allows callers to:
* Get a serializable IOStatisticsSnapshot from an IOStatisticsSource or
IOStatistics instance
* Save an IOStatisticsSnapshot to file
* Convert an IOStatisticsSnapshot to JSON
* Given an object which may be an IOStatisticsSource, return an object
whose toString() value is a dynamically generated, human readable summary.
This is for logging.
* Separate getters to the different sections of IOStatistics.
* Mean values are returned as a Map.Pair<Long, Long> of (samples, sum)
from which means may be calculated.
There are examples of the dynamic bindings to these classes in:
org.apache.hadoop.io.wrappedio.impl.DynamicWrappedIO
org.apache.hadoop.io.wrappedio.impl.DynamicWrappedStatistics
These use DynMethods and other classes in the package
org.apache.hadoop.util.dynamic which are based on the
Apache Parquet equivalents.
This makes re-implementing these in that library and others
which their own fork of the classes (example: Apache Iceberg)
3. The openFile() option "fs.option.openfile.read.policy" has
added specific file format policies for the core filetypes
* avro
* columnar
* csv
* hbase
* json
* orc
* parquet
S3A chooses the appropriate sequential/random policy as a
A policy `parquet, columnar, vector, random, adaptive` will use the parquet policy for
any filesystem aware of it, falling back to the first entry in the list which
the specific version of the filesystem recognizes
4. New Path capability fs.capability.virtual.block.locations
Indicates that locations are generated client side
and don't refer to real hosts.
Contributed by Steve Loughran
This is a followup to #6543 which ensures all test pass in configurations where
fs.s3a.performance.flags is set to "*" or contains "mkdirs"
Contributed by VJ Jasani
If the flag list in fs.s3a.performance.flags includes "mkdir"
then the safety check of a walk up the tree to look for a parent directory,
-done to verify a directory isn't being created under a file- are skipped.
This saves the cost of multiple list operations.
Contributed by Viraj Jasani
1. Configuration adds new method `getEnumSet()` to get a set of enum values from
a configuration string.
<E extends Enum<E>> EnumSet<E> getEnumSet(String key, Class<E> enumClass, boolean ignoreUnknown)
Whitespace is ignored, case is ignored and the value "*" is mapped to "all values of the enum".
If "ignoreUnknown" is true then when parsing, unknown values are ignored.
This is recommended for forward compatiblity with later versions.
2. This support is implemented in org.apache.hadoop.fs.s3a.impl.ConfigurationHelper -it can be used
elsewhere in the hadoop codebase.
3. A new private FlagSet class in hadoop common manages a set of enum flags.
It implements StreamCapabilities and can be probed for a specific option being set
(with a prefix)
S3A adds an option fs.s3a.performance.flags which builds a FlagSet with enum
type PerformanceFlagEnum
* which initially contains {Create, Delete, Mkdir, Open}
* the existing fs.s3a.create.performance option sets the flag "Create".
* tests which configure fs.s3a.create.performance MUST clear
fs.s3a.performance.flags in test setup.
Future performance flags are planned, with different levels of safety
and/or backwards compatibility.
Contributed by Steve Loughran
The new property fs.s3a.encryption.context allow users to specify the AWS KMS Encryption Context to be used in S3A.
The value of the encryption context is a key/value string that will be Base64 encoded and set in the parameter ssekmsEncryptionContext from the S3 client.
Contributed by Raphael Azzolini
Adds new ClientManager interface/implementation which provides on-demand
creation of synchronous and asynchronous s3 clients, s3 transfer manager,
and in close() terminates these.
S3A FS is modified to
* Create a ClientManagerImpl instance and pass down to its S3Store.
* Use the same ClientManager interface against S3Store to demand-create
the services.
* Only create the async client as part of the transfer manager creation,
which will take place during the first rename() operation.
* Statistics on client creation count and duration are recorded.
+ Statistics on the time to initialize and shutdown the S3A FS are collected
in IOStatistics for reporting.
Adds to hadoop common class
LazyAtomicReference<T> implements CallableRaisingIOE<T>, Supplier<T>
and subclass
LazyAutoCloseableReference<T extends AutoCloseable>
extends LazyAtomicReference<T> implements AutoCloseable
These evaluate the Supplier<T>/CallableRaisingIOE<T> they were
constructed with on the first (successful) read of the the value.
Any exception raised during this operation will be rethrown, and on future
evaluations the same operation retried.
These classes implement the Supplier and CallableRaisingIOE
interfaces so can actually be used for to implement lazy function evaluation
as Haskell and some other functional languages do.
LazyAutoCloseableReference is AutoCloseable; its close() method will
close the inner reference if it is set
This class is used in ClientManagerImpl for the lazy S3 Cliehnt creation
and closure.
Contributed by Steve Loughran.
Speed up slow tests
* TestS3AAWSCredentialsProvider: decrease thread pool shutdown time
* TestS3AInputStreamRetry: reduce retry limit and intervals
Contributed by Steve Loughran
The new test TestAWSV2SDK scans the aws sdk bundle.jar and prints out all classes
which are unshaded, so at risk of creating classpath problems
It does not fail the test if this holds, because the current SDKs
do ship with unshaded classes; the test would always fail.
The SDK upgrade process should include inspecting the output
of this test to see if it has got worse (do a before/after check).
Once the AWS SDK does shade everything, we can have this
test fail on any regression
Contributed by Harshit Gupta
It is now possible to provide a job ID in the maven "job.id" property
hadoop-aws test runs to isolate paths under a the test bucket
under which all tests will be executed.
This will allow independent builds *in different source trees*
to test against the same bucket in parallel, and is designed for
CI testing.
Example:
mvn verify -Dparallel-tests -Droot.tests.enabled=false -Djob.id=1
mvn verify -Droot.tests.enabled=false -Djob.id=2
- Root tests must be be disabled to stop them cleaning up
the test paths of other test runs.
- Do still regularly run the root tests just to force cleanup
of the output of any interrupted test suites.
Contributed by Steve Loughran
* parameterize the test run rather than do it from within the test suite.
* log what the committer factory is up to (and improve its logging)
* close all filesystems, then create the test filesystem with cache enabled.
The cache is critical, we want the fs from cache to be used when querying
filesystem properties, rather than one created from the committer jobconf,
which will have the same options as the task committer, so not actually
validate the override logic.
Contributed by Steve Loughran
WrappedIO.bulkDelete_PageSize() => bulkDelete_pageSize()
Makes it consistent with the HADOOP-19131 naming scheme.
The name needs to be fixed before invoking it through reflection,
as once that is attempted the binding won't work at run time,
though compilation will be happy.
Contributed by Steve Loughran
Applications can create a BulkDelete instance from a
BulkDeleteSource; the BulkDelete interface provides
the pageSize(): the maximum number of entries which can be
deleted, and a bulkDelete(Collection paths)
method which can take a collection up to pageSize() long.
This is optimized for object stores with bulk delete APIs;
the S3A connector will offer the page size of
fs.s3a.bulk.delete.page.size unless bulk delete has
been disabled.
Even with a page size of 1, the S3A implementation is
more efficient than delete(path)
as there are no safety checks for the path being a directory
or probes for the need to recreate directories.
The interface BulkDeleteSource is implemented by
all FileSystem implementations, with a page size
of 1 and mapped to delete(pathToDelete, false).
This means that callers do not need to have special
case handling for object stores versus classic filesystems.
To aid use through reflection APIs, the class
org.apache.hadoop.io.wrappedio.WrappedIO
has been created with "reflection friendly" methods.
Contributed by Mukund Thakur and Steve Loughran
HADOOP-19057 switched the hadoop-aws test bucket from landsat-pds to
noaa-cors-pds
This new bucket isn't accessible if the client configuration
sets an fs.s3a.endpoint/region value other than us-east-1.
Contributed by Viraj Jasani
The description of `fs.s3a.committer.abort.pending.uploads` in the section `Concurrent Jobs writing to the same destination` is not correct. Its default value is `true`.
Contributed by Xi Chen
Clarifies behaviour of VectorIO methods with contract tests as well as
specification.
* Add precondition range checks to all implementations
* Identify and fix bug where direct buffer reads was broken
(HADOOP-19101; this surfaced in ABFS contract tests)
* Logging in VectoredReadUtils.
* TestVectoredReadUtils verifies validation logic.
* FileRangeImpl toString() improvements
* CombinedFileRange tracks bytes in range which are wanted;
toString() output logs this.
HDFS
* Add test TestHDFSContractVectoredRead
ABFS
* Add test ITestAbfsFileSystemContractVectoredRead
S3A
* checks for vector IO being stopped in all iterative
vector operations, including draining
* maps read() returning -1 to failure
* passes in file length to validation
* Error reporting to only completeExceptionally() those ranges
which had not yet read data in.
* Improved logging.
readVectored()
* made synchronized. This is only for the invocation;
the actual async retrieves are unsynchronized.
* closes input stream on invocation
* switches to random IO, so avoids keeping any long-lived connection around.
+ AbstractSTestS3AHugeFiles enhancements.
+ ADDENDUM: test fix in ITestS3AContractVectoredRead
Contains: HADOOP-19101. Vectored Read into off-heap buffer broken in fallback
implementation
Contributed by Steve Loughran
Change-Id: Ia4ed71864c595f175c275aad83a2ff5741693432
Clarifies behaviour of VectorIO methods with contract tests as well as specification.
* Add precondition range checks to all implementations
* Identify and fix bug where direct buffer reads was broken
(HADOOP-19101; this surfaced in ABFS contract tests)
* Logging in VectoredReadUtils.
* TestVectoredReadUtils verifies validation logic.
* FileRangeImpl toString() improvements
* CombinedFileRange tracks bytes in range which are wanted;
toString() output logs this.
HDFS
* Add test TestHDFSContractVectoredRead
ABFS
* Add test ITestAbfsFileSystemContractVectoredRead
S3A
* checks for vector IO being stopped in all iterative
vector operations, including draining
* maps read() returning -1 to failure
* passes in file length to validation
* Error reporting to only completeExceptionally() those ranges
which had not yet read data in.
* Improved logging.
readVectored()
* made synchronized. This is only for the invocation;
the actual async retrieves are unsynchronized.
* closes input stream on invocation
* switches to random IO, so avoids keeping any long-lived connection around.
+ AbstractSTestS3AHugeFiles enhancements.
Contains: HADOOP-19101. Vectored Read into off-heap buffer broken in fallback implementation
Contributed by Steve Loughran
If the option fs.s3a.committer.magic.track.commits.in.memory.enabled
is set to true, then rather than save data about in-progress uploads
to S3, this information is cached in memory.
If the number of files being committed is low, this will save network IO
in both the generation of .pending and marker files, and in the scanning
of task attempt directory trees during task commit.
Contributed by Syed Shameerur Rahman
This is a followup to #6406:
HADOOP-18980. S3A credential provider remapping: make extensible
It adds extra validation of key-value pairs in a configuration
option, with tests.
Contributed by Viraj Jasani
FIPS is only supported in north america AWS regions; relevant tests in
ITestS3AEndpointRegion are skipped for buckets with different endpoints/regions.
Disables the new tests added in:
HADOOP-19027. S3A: S3AInputStream doesn't recover from HTTP/channel exceptions #6425
The underlying issue here is that the block prefetch code can identify
when there's a mismatch between declared and actual length, and doesn't
store any of the incomplete buffer.
This should be addressed in HADOOP-18184.
Contributed by Steve Loughran
The AWS landsat data previously used in some S3A tests is no
longer accessible
This PR moves to the new external file
s3a://noaa-cors-pds/raw/2024/001/akse/AKSE001x.24_.gz
* Large enough file for scale tests
* Bucket supports anonymous access
* Ends in .gz to keep codec tests happy
* No spaces in path to keep bucket-info happy
Test Code Changes
* Leaves the test key name alone: fs.s3a.scale.test.csvfile
* Rename all methods and fields move remove "csv" from their names and
move to "external file" we no longer require it to be CSV.
* Path definition and helper methods have been moved to PublicDatasetTestUtils
* Improve error reporting in ITestS3AInputStreamPerformance if the file
is too short
With S3 Select removed, there is no need for the file to be
a CSV file; there is a test which tries to unzip it; other
tests have a minimum file size.
Consult the JIRA for the settings to add to auth-keys.xml
to switch earlier builds to this same file.
Contributed by Steve Loughran
This is a followup to PR:
HADOOP-19045. S3A: Validate CreateSession Timeout Propagation (#6470)
Remove all declarations of fs.s3a.connection.request.timeout
in
- hadoop-common/src/main/resources/core-default.xml
- hadoop-aws/src/test/resources/core-site.xml
New test in TestAwsClientConfig to verify that the value
defined in fs.s3a.Constants class is used.
This is brittle to someone overriding it in their test setups,
but as this test is intended to verify that the option is not
explicitly set, there's no workaround.
Contributed by Steve Loughran
The option fs.s3a.classloader.isolation (default: true) can be set to false to disable s3a classloader isolation;
This can assist in using custom credential providers and other extension points.
Contributed by Antonio Murgia
Improves region handling in the S3A connector, including enabling cross-region support
when that is considered necessary.
Consult the documentation in connecting.md/connecting.html for the current
resolution process.
Contributed by Viraj Jasani
Cut out S3 Select
* leave public/unstable constants alone
* s3guard tool will fail with error
* s3afs. path capability will fail
* openFile() will fail with specific error
* s3 select doc updated
* Cut eventstream jar
* New test: ITestSelectUnsupported verifies new failure
handling above
Contributed by Steve Loughran
New test ITestCreateSessionTimeout to verify that the duration set
in fs.s3a.connection.request.timeout is passed all the way down.
This is done by adding a sleep() in a custom signer and verifying
that it is interrupted and that an AWSApiCallTimeoutException is
raised.
+ Fix testRequestTimeout()
* doesn't skip if considered cross-region
* sets a minimum duration of 0 before invocation
* resets the minimum afterwards
Contributed by Steve Loughran
This update ensures that the timeout set in fs.s3a.connection.request.timeout is passed down
to calls to CreateSession made in the AWS SDK to get S3 Express session tokens.
Contributed by Steve Loughran
Add new option fs.s3a.checksum.validation, default false, which
is used when creating s3 clients to enable/disable checksum
validation.
When false, GET response processing is measurably faster.
Contributed by Steve Loughran.
HADOOP-19015. Increase fs.s3a.connection.maximum to 500 to minimize the risk of Timeout waiting for connection from the pool
Contributed By: Mukund Thakur
Adds a new option `fs.s3a.endpoint.fips` to switch the SDK client to use
FIPS endpoints, as an alternative to explicitly declaring them.
* The option is available as a path capability for probes.
* SDK v2 itself doesn't know that some regions don't have FIPS endpoints
* SDK only fails with endpoint + fips flag as a retried exception; wit this
change the S3A client should fail fast.
PR fails fast.
* Adds a new "connecting.md" doc; moves existing docs there and restructures.
* New Tests in ITestS3AEndpointRegion
bucket-info command support:
* added to list of path capabilities
* added -fips flag and test for explicit probe
* also now prints bucket region
* and removed some of the obsolete s3guard options
* updated docs
Contributed by Steve Loughran
Differentiate from "EOF out of range/end of GET" from
"EOF channel problems" through
two different subclasses of EOFException and input streams to always
retry on http channel errors; out of range GET requests are not retried.
Currently an EOFException is always treated as a fail-fast call in read()
This allows for all existing external code catching EOFException to handle
both, but S3AInputStream to cleanly differentiate range errors (map to -1)
from channel errors (retry)
- HttpChannelEOFException is subclass of EOFException, so all code
which catches EOFException is still happy.
retry policy: connectivityFailure
- RangeNotSatisfiableEOFException is the subclass of EOFException
raised on 416 GET range errors.
retry policy: fail
- Method ErrorTranslation.maybeExtractChannelException() to create this
from shaded/unshaded NoHttpResponseException, using string match to
avoid classpath problems.
- And do this for SdkClientExceptions with OpenSSL error code WFOPENSSL0035.
We believe this is the OpenSSL equivalent.
- ErrorTranslation.maybeExtractIOException() to perform this translation as
appropriate.
S3AInputStream.reopen() code retries on EOF, except on
RangeNotSatisfiableEOFException,
which is converted to a -1 response to the caller
as is done historically.
S3AInputStream knows to handle these with
read(): HttpChannelEOFException: stream aborting close then retry
lazySeek(): Map RangeNotSatisfiableEOFException to -1, but do not map
any other EOFException class raised.
This means that
* out of range reads map to -1
* channel problems in reopen are retried
* channel problems in read() abort the failed http connection so it
isn't recycled
Tests for this using/abusing mocking.
Testing through actually raising 416 exceptions and verifying that
readFully(), char read() and vector reads are all good.
There is no attempt to recover within a readFully(); there's
a boolean constant switch to turn this on, but if anyone does
it a test will spin forever as the inner PositionedReadable.read(position, buffer, len)
downgrades all EOF exceptions to -1.
A new method would need to be added which controls whether to downgrade/rethrow
exceptions.
What does that mean? Possibly reduced resilience to non-retried failures
on the inner stream, even though more channel exceptions are retried on.
Contributed by Steve Loughran
