[
https://issues.apache.org/jira/browse/HADOOP-14457?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16049856#comment-16049856
]
Aaron Fabbri commented on HADOOP-14457:
---------------------------------------
Good summary [~mackrorysd]. One thing I'll add about my preference here (which
is to commit your patch and add a capabilities check as a performance
optimization, pending fixing create()), is the design of the MetadataStore
interface. I prefer it to be:
1. Simple to implement.
2. Not specific to S3A internals.
3. Explicit in the way state is mutated (a trailing log of FS metadata changes).
For #3, I prefer the basic contract to remain that an FS client reports
metadata changes it makes explicitly. When create() happens, for example, if
the FS client is creating ancestor directories, I'd like to see those creations
reported directly via the API, instead of requiring all implementations to
deduce them from the path. Besides the design preference, *consuming the
stream of modifications as they happen also makes things like pruning or TTL
expiry possible* as we see directory creation when it happens. We have no idea
of the age of a directory path, for instance, if we allow the FS to skip
reporting its creation.
In my experience explicit APIs like this are easier to reason about and less
likely to hide bugs and allow for general sloppiness.
Finally, I do not think this case of create() being broken is compelling enough
to change the above design that exists today. I also don't feel like it
warrants the additional complexity of a putRecursive() method on MetadataStore
(which also suffers from the TTL issue). Separation of concerns here, to me,
means S3A's broken create() does not infect MetadataStore.
> create() does not notify metadataStore of parent directories or ensure
> they're not existing files
> -------------------------------------------------------------------------------------------------
>
> Key: HADOOP-14457
> URL: https://issues.apache.org/jira/browse/HADOOP-14457
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Reporter: Sean Mackrory
> Assignee: Sean Mackrory
> Attachments: HADOOP-14457-HADOOP-13345.001.patch,
> HADOOP-14457-HADOOP-13345.002.patch, HADOOP-14457-HADOOP-13345.003.patch,
> HADOOP-14457-HADOOP-13345.004.patch, HADOOP-14457-HADOOP-13345.005.patch,
> HADOOP-14457-HADOOP-13345.006.patch, HADOOP-14457-HADOOP-13345.007.patch,
> HADOOP-14457-HADOOP-13345.008.patch, HADOOP-14457-HADOOP-13345.009.patch
>
>
> Not a great test yet, but it at least reliably demonstrates the issue.
> LocalMetadataStore will sometimes erroneously report that a directory is
> empty with isAuthoritative = true when it *definitely* has children the
> metadatastore should know about. It doesn't appear to happen if the children
> are just directory. The fact that it's returning an empty listing is
> concerning, but the fact that it says it's authoritative *might* be a second
> bug.
> {code}
> diff --git
> a/hadoop-tools/hadoop-aws/src/main/java/org/apache/hadoop/fs/s3a/S3AFileSystem.java
>
> b/hadoop-tools/hadoop-aws/src/main/java/org/apache/hadoop/fs/s3a/S3AFileSystem.java
> index 78b3970..1821d19 100644
> ---
> a/hadoop-tools/hadoop-aws/src/main/java/org/apache/hadoop/fs/s3a/S3AFileSystem.java
> +++
> b/hadoop-tools/hadoop-aws/src/main/java/org/apache/hadoop/fs/s3a/S3AFileSystem.java
> @@ -965,7 +965,7 @@ public boolean hasMetadataStore() {
> }
>
> @VisibleForTesting
> - MetadataStore getMetadataStore() {
> + public MetadataStore getMetadataStore() {
> return metadataStore;
> }
>
> diff --git
> a/hadoop-tools/hadoop-aws/src/test/java/org/apache/hadoop/fs/contract/s3a/ITestS3AContractRename.java
>
> b/hadoop-tools/hadoop-aws/src/test/java/org/apache/hadoop/fs/contract/s3a/ITestS3AContractRename.java
> index 4339649..881bdc9 100644
> ---
> a/hadoop-tools/hadoop-aws/src/test/java/org/apache/hadoop/fs/contract/s3a/ITestS3AContractRename.java
> +++
> b/hadoop-tools/hadoop-aws/src/test/java/org/apache/hadoop/fs/contract/s3a/ITestS3AContractRename.java
> @@ -23,6 +23,11 @@
> import org.apache.hadoop.fs.contract.AbstractFSContract;
> import org.apache.hadoop.fs.FileSystem;
> import org.apache.hadoop.fs.Path;
> +import org.apache.hadoop.fs.s3a.S3AFileSystem;
> +import org.apache.hadoop.fs.s3a.Tristate;
> +import org.apache.hadoop.fs.s3a.s3guard.DirListingMetadata;
> +import org.apache.hadoop.fs.s3a.s3guard.MetadataStore;
> +import org.junit.Test;
>
> import static org.apache.hadoop.fs.contract.ContractTestUtils.dataset;
> import static org.apache.hadoop.fs.contract.ContractTestUtils.writeDataset;
> @@ -72,4 +77,24 @@ public void testRenameDirIntoExistingDir() throws
> Throwable {
> boolean rename = fs.rename(srcDir, destDir);
> assertFalse("s3a doesn't support rename to non-empty directory", rename);
> }
> +
> + @Test
> + public void testMkdirPopulatesFileAncestors() throws Exception {
> + final FileSystem fs = getFileSystem();
> + final MetadataStore ms = ((S3AFileSystem) fs).getMetadataStore();
> + final Path parent = path("testMkdirPopulatesFileAncestors/source");
> + try {
> + fs.mkdirs(parent);
> + final Path nestedFile = new Path(parent, "dir1/dir2/dir3/file4");
> + byte[] srcDataset = dataset(256, 'a', 'z');
> + writeDataset(fs, nestedFile, srcDataset, srcDataset.length,
> + 1024, false);
> +
> + DirListingMetadata list = ms.listChildren(parent);
> + assertTrue("MetadataStore falsely reports authoritative empty list",
> + list.isEmpty() == Tristate.FALSE || !list.isAuthoritative());
> + } finally {
> + fs.delete(parent, true);
> + }
> + }
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
