org.apache.hadoop.fs.s3a.s3guard

Class DynamoDBMetadataStore

java.lang.Object

org.apache.hadoop.fs.s3a.s3guard.DynamoDBMetadataStore

All Implemented Interfaces:

Closeable, AutoCloseable, MetadataStore

@InterfaceAudience.Private
@InterfaceStability.Evolving
public class DynamoDBMetadataStore
extends Object
implements MetadataStore

DynamoDBMetadataStore is a MetadataStore that persists file system metadata to DynamoDB. The current implementation uses a schema consisting of a single table. The name of the table can be configured by config key Constants.S3GUARD_DDB_TABLE_NAME_KEY. By default, it matches the name of the S3 bucket. Each item in the table represents a single directory or file. Its path is split into separate table attributes:

• parent (absolute path of the parent, with bucket name inserted as first path component).
• child (path of that specific child, relative to parent).
• optional boolean attribute tracking whether the path is a directory. Absence or a false value indicates the path is a file.
• optional long attribute revealing modification time of file. This attribute is meaningful only to file items.
• optional long attribute revealing file length. This attribute is meaningful only to file items.
• optional long attribute revealing block size of the file. This attribute is meaningful only to file items.

The DynamoDB partition key is the parent, and the range key is the child. To allow multiple buckets to share the same DynamoDB table, the bucket name is treated as the root directory. For example, assume the consistent store contains metadata representing this file system structure:

 s3a://bucket/dir1
 |-- dir2
 |   |-- file1
 |   `-- file2
 `-- dir3
     |-- dir4
     |   `-- file3
     |-- dir5
     |   `-- file4
     `-- dir6
 

This is persisted to a single DynamoDB table as:

 =========================================================================
 | parent                 | child | is_dir | mod_time | len |     ...    |
 =========================================================================
 | /bucket                | dir1  | true   |          |     |            |
 | /bucket/dir1           | dir2  | true   |          |     |            |
 | /bucket/dir1           | dir3  | true   |          |     |            |
 | /bucket/dir1/dir2      | file1 |        |   100    | 111 |            |
 | /bucket/dir1/dir2      | file2 |        |   200    | 222 |            |
 | /bucket/dir1/dir3      | dir4  | true   |          |     |            |
 | /bucket/dir1/dir3      | dir5  | true   |          |     |            |
 | /bucket/dir1/dir3/dir4 | file3 |        |   300    | 333 |            |
 | /bucket/dir1/dir3/dir5 | file4 |        |   400    | 444 |            |
 | /bucket/dir1/dir3      | dir6  | true   |          |     |            |
 =========================================================================
 

This choice of schema is efficient for read access patterns. get(Path) can be served from a single item lookup. listChildren(Path) can be served from a query against all rows matching the parent (the partition key) and the returned list is guaranteed to be sorted by child (the range key). Tracking whether or not a path is a directory helps prevent unnecessary queries during traversal of an entire sub-tree. Some mutating operations, notably deleteSubtree(Path) and move(Collection, Collection), are less efficient with this schema. They require mutating multiple items in the DynamoDB table. By default, DynamoDB access is performed within the same AWS region as the S3 bucket that hosts the S3A instance. During initialization, it checks the location of the S3 bucket and creates a DynamoDB client connected to the same region. The region may also be set explicitly by setting the config parameter fs.s3a.s3guard.ddb.region to the corresponding region.

Field Summary

Fields

Modifier and Type	Field and Description
static String	E_INCOMPATIBLE_VERSION Error: version mismatch.
static String	E_NO_VERSION_MARKER Error: version marker not found in table.
static org.slf4j.Logger	LOG
static long	MIN_RETRY_SLEEP_MSEC Initial delay for retries when batched operations get throttled by DynamoDB.
static int	VERSION Current version number.
static String	VERSION_MARKER parent/child name to use in the version marker.

Constructor Summary

Constructors

Constructor and Description
DynamoDBMetadataStore()

Method Summary

Methods

Modifier and Type	Method and Description
void	close()
void	delete(org.apache.hadoop.fs.Path path) Deletes exactly one path, leaving a tombstone to prevent lingering, inconsistent copies of it from being listed.
void	deleteSubtree(org.apache.hadoop.fs.Path path) Deletes the entire sub-tree rooted at the given path, leaving tombstones to prevent lingering, inconsistent copies of it from being listed.
void	destroy() Destroy all resources associated with the metadata store.
void	forgetMetadata(org.apache.hadoop.fs.Path path) Removes the record of exactly one path.
PathMetadata	get(org.apache.hadoop.fs.Path path) Gets metadata for a path.
PathMetadata	get(org.apache.hadoop.fs.Path path, boolean wantEmptyDirectoryFlag) Gets metadata for a path.
Map<String,String>	getDiagnostics() Get any diagnostics information from a store, as a list of (key, value) tuples for display.
void	initialize(org.apache.hadoop.conf.Configuration config) Performs one-time initialization of the metadata store via configuration.
void	initialize(org.apache.hadoop.fs.FileSystem fs) Performs one-time initialization of the metadata store.
DirListingMetadata	listChildren(org.apache.hadoop.fs.Path path) Lists metadata for all direct children of a path.
void	move(Collection<org.apache.hadoop.fs.Path> pathsToDelete, Collection<PathMetadata> pathsToCreate) Record the effects of a FileSystem.rename(Path, Path) in the MetadataStore.
void	prune(long modTime) Clear any metadata older than a specified time from the repository.
void	put(Collection<PathMetadata> metas) Saves metadata for any number of paths.
void	put(DirListingMetadata meta) Save directory listing metadata.
void	put(PathMetadata meta) Saves metadata for exactly one path.
String	toString()
void	updateParameters(Map<String,String> parameters) Tune/update parameters for an existing table.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

LOG

public static final org.slf4j.Logger LOG

VERSION_MARKER

public static final String VERSION_MARKER

parent/child name to use in the version marker.

See Also:

Constant Field Values

VERSION

public static final int VERSION

Current version number.

See Also:

Constant Field Values

E_NO_VERSION_MARKER

public static final String E_NO_VERSION_MARKER

Error: version marker not found in table.

See Also:

Constant Field Values

E_INCOMPATIBLE_VERSION

public static final String E_INCOMPATIBLE_VERSION

Error: version mismatch.

See Also:

Constant Field Values

MIN_RETRY_SLEEP_MSEC

public static final long MIN_RETRY_SLEEP_MSEC

Initial delay for retries when batched operations get throttled by DynamoDB. Value is 100L msec.

See Also:

Constant Field Values

Constructor Detail

DynamoDBMetadataStore

public DynamoDBMetadataStore()

Method Detail

initialize

public void initialize(org.apache.hadoop.fs.FileSystem fs)
                throws IOException

Description copied from interface: MetadataStore

Performs one-time initialization of the metadata store.

Specified by:

initialize in interface MetadataStore

Parameters:

fs - FileSystem associated with the MetadataStore

Throws:

IOException - if there is an error

initialize

public void initialize(org.apache.hadoop.conf.Configuration config)
                throws IOException

Performs one-time initialization of the metadata store via configuration. This initialization depends on the configuration object to get AWS credentials, DynamoDBFactory implementation class, DynamoDB endpoints, DynamoDB table names etc. After initialization, this metadata store does not explicitly relate to any S3 bucket, which be nonexistent. This is used to operate the metadata store directly beyond the scope of the S3AFileSystem integration, e.g. command line tools. Generally, callers should use initialize(FileSystem) with an initialized S3AFileSystem instance. Without a filesystem to act as a reference point, the configuration itself must declare the table name and region in the Constants.S3GUARD_DDB_TABLE_NAME_KEY and Constants.S3GUARD_DDB_REGION_KEY respectively.

Specified by:

initialize in interface MetadataStore

Parameters:

config - Configuration.

Throws:

IOException - if there is an error

IllegalArgumentException - if the configuration is incomplete

See Also:

initialize(FileSystem)

delete

public void delete(org.apache.hadoop.fs.Path path)
            throws IOException

Description copied from interface: MetadataStore

Deletes exactly one path, leaving a tombstone to prevent lingering, inconsistent copies of it from being listed.

Specified by:

delete in interface MetadataStore

Parameters:

path - the path to delete

Throws:

IOException - if there is an error

forgetMetadata

public void forgetMetadata(org.apache.hadoop.fs.Path path)
                    throws IOException

Description copied from interface: MetadataStore

Removes the record of exactly one path. Does not leave a tombstone (see MetadataStore.delete(Path). It is currently intended for testing only, and a need to use it as part of normal FileSystem usage is not anticipated.

Specified by:

forgetMetadata in interface MetadataStore

Parameters:

path - the path to delete

Throws:

IOException - if there is an error

deleteSubtree

public void deleteSubtree(org.apache.hadoop.fs.Path path)
                   throws IOException

Description copied from interface: MetadataStore

Deletes the entire sub-tree rooted at the given path, leaving tombstones to prevent lingering, inconsistent copies of it from being listed. In addition to affecting future calls to MetadataStore.get(Path), implementations must also update any stored DirListingMetadata objects which track the parent of this file.

Specified by:

deleteSubtree in interface MetadataStore

Parameters:

path - the root of the sub-tree to delete

Throws:

IOException - if there is an error

get

public PathMetadata get(org.apache.hadoop.fs.Path path)
                 throws IOException

Description copied from interface: MetadataStore

Gets metadata for a path.

Specified by:

get in interface MetadataStore

Parameters:

path - the path to get

Returns:

metadata for path, null if not found

Throws:

IOException - if there is an error

get

public PathMetadata get(org.apache.hadoop.fs.Path path,
               boolean wantEmptyDirectoryFlag)
                 throws IOException

Description copied from interface: MetadataStore

Gets metadata for a path. Alternate method that includes a hint whether or not the MetadataStore should do work to compute the value for PathMetadata.isEmptyDirectory(). Since determining emptiness may be an expensive operation, this can save wasted work.

Specified by:

get in interface MetadataStore

Parameters:

path - the path to get

wantEmptyDirectoryFlag - Set to true to give a hint to the MetadataStore that it should try to compute the empty directory flag.

Returns:

metadata for path, null if not found

Throws:

IOException - if there is an error

listChildren

public DirListingMetadata listChildren(org.apache.hadoop.fs.Path path)
                                throws IOException

Description copied from interface: MetadataStore

Lists metadata for all direct children of a path.

Specified by:

listChildren in interface MetadataStore

Parameters:

path - the path to list

Returns:

metadata for all direct children of path which are being tracked by the MetadataStore, or null if the path was not found in the MetadataStore.

Throws:

IOException - if there is an error

move

public void move(Collection<org.apache.hadoop.fs.Path> pathsToDelete,
        Collection<PathMetadata> pathsToCreate)
          throws IOException

Description copied from interface: MetadataStore

Record the effects of a FileSystem.rename(Path, Path) in the MetadataStore. Clients provide explicit enumeration of the affected paths (recursively), before and after the rename. This operation is not atomic, unless specific implementations claim otherwise. On the need to provide an enumeration of directory trees instead of just source and destination paths: Since a MetadataStore does not have to track all metadata for the underlying storage system, and a new MetadataStore may be created on an existing underlying filesystem, this move() may be the first time the MetadataStore sees the affected paths. Therefore, simply providing src and destination paths may not be enough to record the deletions (under src path) and creations (at destination) that are happening during the rename().

Specified by:

move in interface MetadataStore

Parameters:

pathsToDelete - Collection of all paths that were removed from the source directory tree of the move.

pathsToCreate - Collection of all PathMetadata for the new paths that were created at the destination of the rename ().

Throws:

IOException - if there is an error

put

public void put(PathMetadata meta)
         throws IOException

Description copied from interface: MetadataStore

Saves metadata for exactly one path. Implementations may pre-create all the path's ancestors automatically. Implementations must update any DirListingMetadata objects which track the immediate parent of this file.

Specified by:

put in interface MetadataStore

Parameters:

meta - the metadata to save

Throws:

IOException - if there is an error

put

public void put(Collection<PathMetadata> metas)
         throws IOException

Description copied from interface: MetadataStore

Saves metadata for any number of paths. Semantics are otherwise the same as single-path puts.

Specified by:

put in interface MetadataStore

Parameters:

metas - the metadata to save

Throws:

IOException - if there is an error

put

public void put(DirListingMetadata meta)
         throws IOException

Description copied from interface: MetadataStore

Save directory listing metadata. Callers may save a partial directory listing for a given path, or may store a complete and authoritative copy of the directory listing. MetadataStore implementations may subsequently keep track of all modifications to the directory contents at this path, and return authoritative results from subsequent calls to MetadataStore.listChildren(Path). See DirListingMetadata. Any authoritative results returned are only authoritative for the scope of the MetadataStore: A per-process MetadataStore, for example, would only show results visible to that process, potentially missing metadata updates (create, delete) made to the same path by another process.

Specified by:

put in interface MetadataStore

Parameters:

meta - Directory listing metadata.

Throws:

IOException - if there is an error

close

public void close()

Specified by:

close in interface Closeable

Specified by:

close in interface AutoCloseable

destroy

public void destroy()
             throws IOException

Description copied from interface: MetadataStore

Destroy all resources associated with the metadata store. The destroyed resources can be DynamoDB tables, MySQL databases/tables, or HDFS directories. Any operations after calling this method may possibly fail. This operation is idempotent.

Specified by:

destroy in interface MetadataStore

Throws:

IOException - if there is an error

prune

public void prune(long modTime)
           throws IOException

Description copied from interface: MetadataStore

Clear any metadata older than a specified time from the repository. Implementations MUST clear file metadata, and MAY clear directory metadata (s3a itself does not track modification time for directories). Implementations may also choose to throw UnsupportedOperationException istead. Note that modification times should be in UTC, as returned by System.currentTimeMillis at the time of modification.

Specified by:

prune in interface MetadataStore

Parameters:

modTime - Oldest modification time to allow

Throws:

IOException - if there is an error

toString

public String toString()

Overrides:

toString in class Object

getDiagnostics

public Map<String,String> getDiagnostics()
                                  throws IOException

Description copied from interface: MetadataStore

Get any diagnostics information from a store, as a list of (key, value) tuples for display. Arbitrary values; no guarantee of stability. These are for debugging only.

Specified by:

getDiagnostics in interface MetadataStore

Returns:

a map of strings.

Throws:

IOException - if there is an error

updateParameters

public void updateParameters(Map<String,String> parameters)
                      throws IOException

Description copied from interface: MetadataStore

Tune/update parameters for an existing table.

Specified by:

updateParameters in interface MetadataStore

Parameters:

parameters - map of params to change.

Throws:

IOException - if there is an error