Fork me on GitHub

The Blob Store

The Oak BlobStore is similar to the Jackrabbit 2.x DataStore. However, there are a few minor problems the BlobStore tries to address. Because, for the Jackrabbit DataStore:

  • a temporary file is created when adding a large binary, even if the binary already exists

  • sharding is slow and complicated because the hash needs to be calculated first, before the binary is stored in the target shard (the FileDataStore still doesn't support sharding the directory currently)

  • file handles are kept open until the consumer is done reading, which complicates the code, and we could potentially get “too many open files” when the consumer doesn't close the stream

  • for database based data stores, there is a similar (even worse) problem that streams are kept open, which means we need to use connection pooling, and if the user doesn't close the stream we could run out of connections

  • for database based data stores, for some databases (MySQL), binaries are fully read in memory, which results in out-of-memory

  • binaries that are similar are always stored separately no matter what

Those problems are solved in Oak BlobStores, because binaries are split into blocks of 2 MB. This is similar to how DropBox works internally. Blocks are processed in memory so that temp files are never needed, and blocks are cached. File handles don't need to be kept open. Sharding is trivial because each block is processed separately.

Binaries that are similar: in the BlobStore, currently, they are stored separately except if some of the 2 MB blocks match. However, the algorithm in the BlobStore would allow to re-use all matching parts, because in the BlobStore, concatenating blob ids means concatenating the data.

Another change was that most DataStore implementations use SHA-1, while the BlobStore uses SHA-256. Using SHA-256 will be a requirement at some point, see also http://en.wikipedia.org/wiki/SHA-2 “Federal agencies … must use the SHA-2 family of hash functions for these applications after 2010”. This might affect some potential users.

Support for Jackrabbit 2 DataStore

Jackrabbit 2 used DataStore to store blobs. Oak supports usage of such DataStore via DataStoreBlobStore wrapper. This allows usage of FileDataStore and S3DataStore with Oak NodeStore implementations.

NodeStore and BlobStore

Currently Oak provides two NodeStore implementations i.e. SegmentNodeStore and DocumentNodeStore. Further Oak ships with multiple BlobStore implementations

  1. FileBlobStore - Stores the file contents in chunks on file system
  2. MongoBlobStore - Stores the file content in chunks in Mongo. Typically used with DocumentNodeStore when running on Mongo by default
  3. FileDataStore (with wrapper) - Stores the file on file system without breaking it into chunks. Mostly used when blobs have to shared between multiple repositories. Also used by default when migrating Jackrabbit 2 repositories to Oak
  4. S3DataStore (with wrapper) - Stores the file in Amazon S3
  5. RDBBlobStore - Store the file contents in chunks in a relational databases. Typically used with DocumentNodeStorewhen using a relational DB persistence
  6. AzureDataStore (with wrapper) - Stores the file in Microsoft Azure Blob storage

In addition there are some more implementations which are considered experimental

  1. CloudBlobStore - Stores the file file chunks in cloud storage using the JClouds BlobStore API.
  2. MongoGridFSBlobStore - Stores the file chunks in Mongo using GridFS support

Depending on NodeStore type and usage requirement these can be configured to use a particular BlobStore implementation. For OSGi env refer to Configuring DataStore/BlobStore

SegmentNodeStore (TarMK)

By default SegmentNodeStore (aka TarMK) does not require a BlobStore. Instead the binary content is directly stored as part of segment blob itself. Depending on requirements one of the following can be used

  • FileDataStore - This should be used if the blobs/binaries have to be shared between multiple repositories. This would also be used when a JR2 repository is migrated to Oak
  • S3DataStore - This should be used when binaries are stored in Amazon S3
  • AzureDataStore - This should be used when binaries are stored in Microsoft Azure Blob storage

DocumentNodeStore

By default DocumentNodeStore when running on Mongo uses MongoBlobStore. Depending on requirements one of the following can be used

  • MongoBlobStore - Used by default and recommended only for development and testing.
  • FileDataStore - This should be used if the binaries have to be stored on the file system. This would also be used when a JR2 repository is migrated to Oak
  • S3DataStore - This should be used when binaries are stored in Amazon S3. Typically used when running in Amazon AWS
  • AzureDataStore - This should be used when binaries are stored in Microsoft Azure Blob storage

Caching DataStore

The DataStore implementations S3DataStore,CachingFileDataStore and AzureDataStore support local file system caching for the files/blobs and extend the AbstractSharedCachingDataStore class which implements the caching functionality. The CachingFileDataStore is useful when the DataStore is on nfs. The cache has a size limit and is configured by the cacheSize parameter.

Downloads

The local cache will be checked for existence of the record corresponding to the requested file/blob before accessing it from the DataStore. When the cache exceeds the limit configured while adding a file into the cache then some of the file(s) will be evicted to reclaim space.

Asynchronous Uploads

The cache also supports asynchronous uploads to the DataStore. The files are staged locally in the cache on the file system and an asynchronous job started to upload the file. The number of asynchronous uploads are limited by the size of staging cache configured by the stagingSplitPercentage parameter and is by default set to 10. This defines the ratio of the cacheSize to be dedicated for the staging cache. The percentage of cache available for downloads is calculated as (100 - stagingSplitPerentage) * cacheSize (by default 90). The asynchronous uploads are also multi-threaded and is governed by the uploadThreads configuration parameter. The default value is 10.

The files are moved to the main download cache after the uploads are complete. When the staging cache exceeds the limit, the files are uploaded synchronously to the DataStore until the previous asynchronous uploads are complete and space available in the staging cache. The uploaded files are removed from the staging area by a periodic job whose interval is configured by the stagingPurgeInterval configuration parameter. The default value is 300 seconds.

Any failed uploads (due to various reasons e.g. network disruption) are put on a retry queue and retried periodically with the configured interval stagingRetryInterval. The default value for is 600 seconds.

Caching Stats

The ConsolidatedDataStoreCacheStats is registered as an MBean and provides a snapshot of the cache performance for both the download and the upload staging cache.

datastore cache stats

The following table explains the different statistics exposed for both type of caches

Parameters DataStore-DownloadCache DataStore-StagingCache
elementCount Number of files cached Pending file uploads in cache
requestCount Number of files requested from cache Number of file uploads requested
hitCount Number of files served from cache Number of files uploaded asynchronously
hitRate Ratio of hits to requests Ratio of hits to requests
loadCount Number of files loaded when not in cache Number of file requests from cache
loadSuccessCount Number of files successfully loaded Number of file requests served from cache
loadExceptionCount Number of load file unsuccessful Number of file requests not in cache
maxWeight Max cache size (bytes) Max cache size (bytes)
totalWeight Current size of cache (bytes Current size of cache (bytes)bytes
totalMemWeight Approximate size of cache in-memory (bytes) Approximate size of cache in memory (bytes)

The parameters above can be used to size the cache. For example:

  • The hitRate is an important parameter and if much below 1 then indicates that the cache is low based on the load and should be increased.
  • If the staging cache has a low hit ratio, download cache has a high hit ratio and also its current size is much less than the maxSize then its better to increase the stagingSplitPercentage parameter.

The MBean also exposes a method isFileSynced which takes a node path of a binary and returns whether the associated file/blob has been uploaded to the DataStore.

Upgrade (Pre Oak 1.6 caching)

When upgrading from the older cache implementation the process should be seamless and any pending uploads would be scheduled for upload and any previously downloaded files in the cache will be put in the cache on initialization. There is a slight difference in the structure of the local file system cache directory. Whereas in the older cache structure both the downloaded and the upload files were put directly under the cache path. The newer structure segregates the downloads and uploads and stores them under cache path under the directories download and upload
respectively.

There is also an option to upgrade the cache offline by using the datastorecacheupgrade command of oak-run. The details on how to execute the command and the different parameters can be checked in the readme for the oak-run module.

Blob Garbage Collection

Blob Garbage Collection(GC) is applicable for the following blob stores:

  • DocumentNodeStore

    • MongoBlobStore/RDBBlobStore (Default blob stores for RDB & Mongo)
    • FileDataStore
    • S3DataStore
    • SharedS3DataStore (since Oak 1.2.0)
    • AzureDataStore
  • SegmentNodeStore

    • FileDataStore
    • S3DataStore
    • SharedS3DataStore (since Oak 1.2.0)
    • AzureDataStore

Oak implements a Mark and Sweep based Garbage Collection logic.

  1. Mark Phase - In this phase the binary references are marked in both BlobStore and NodeStore
    1. Mark BlobStore - GC logic would make a record of all the blobs present in the BlobStore.
    2. Mark NodeStore - GC logic would make a record of all the blob references which are referred by any node present in NodeStore. Note that any blob references from old revisions of node would also be considered as a valid references.
  2. Sweep Phase - In this phase all blob references form Mark BlobStore phase which were not found in Mark NodeStore part would considered as GC candidates. It would only delete blobs which are older than a specified time interval (last modified say 24 hrs (default) ago).

The garbage collection can be triggered by calling:

  • MarkSweepGarbageCollector#collectGarbage() (Oak 1.0.x)
  • MarkSweepGarbageCollector#collectGarbage(false) (Oak 1.2.x)
  • If the MBeans are registered in the MBeanServer then the following can also be used to trigger GC:
    • BlobGC#startBlobGC() which takes in a markOnly boolean parameter to indicate mark only or complete gc

Caching of Blob ids locally (Oak 1.6.x)

For the FileDataStore, S3DataStore and AzureDataStore the blob ids are cached locally on the disk when they are created which speeds up the ‘Mark BlobStore’ phase. The locally tracked ids are synchronized with the data store periodically to enable other cluster nodes or different repositories sharing the datastore to get a consolidated list of all blob ids. The interval of synchronization is defined by the OSGi configuration parameter blobTrackSnapshotIntervalInSecs for the configured NodeStore services.

If 2 garbage collection cycles are executed within the blobTrackSnapshotIntervalInSecs then there may be warnings in the logs of some missing blob ids which is due to the fact that the deletions due to earlier gc has not been synchronized with the data store. It's ok to either ignore these warnings or to adjust the blobTrackSnapshotIntervalInSecs parameter according to the schedule identified for running blob gc.

When upgrading an existing system to take advantage of caching the existing blob ids have to be cached. One of the following should be executed.

  • Use MarkSweepGarbageCollectot#collectGarbage(boolean markOnly, boolean forceBlobRetrieve) with true for forceBlobRetrieve parameter to force retrieving blob ids from the datastore and cache locally also.
  • Execute Blob GC before the configured time duration of blobTrackSnapshotIntervalInSecs.
  • Execute consistency check from the JMX BlobGCMbean before the configured time duration of blobTrackSnapshotIntervalInSecs.
  • Execute datastorecheck command offline using oak-run with the --track option as defined in consistency check.

Shared DataStore Blob Garbage Collection (Since 1.2.0)

Registration

On start of a repository configured to use a shared DataStore (same path, S3 bucket or Azure container), a unique repository id is generated and registered in the NodeStore as well as the DataStore. In the DataStore this repository id is registered as an empty file with the format repository-[repository-id] (e.g. repository-988373a0-3efb-451e-ab4c-f7e794189273). This empty file is created under:

  • FileDataStore - Under the root directory configured for the datastore.
  • S3DataStore - Under META folder in the S3 bucket configured.
  • AzureDataStore - Under META folder in the Azure container configured.

On start/configuration of all the repositories sharing the data store it should be confirmed that the unique repositoryId per repository is registered in the DataStore. Refer the section below on Checking Shared GC status.

Execution

The high-level process for garbage collection is still the same as described above. But to support blob garbage collection in a shared DataStore the Mark and Sweep phase can be run independently.

The details of the process are as follows:

  • The Mark NodeStore phase has to be executed for each of the repositories sharing the DataStore.
    • This can be executed by running MarkSweepGarbageCollector#collectGarbage(true), where true indicates mark only.
    • All the references are collected in the DataStore in a file with the format references-[repository-id] (e.g. references-988373a0-3efb-451e-ab4c-f7e794189273).
  • One completion of the above process on all repositories, the sweep phase needs to be triggered.
    • This can be executed by running MarkSweepGarbageCollector#collectGarbage(false) on one of the repositories, where false indicates to run sweep also.
    • The sweep process checks for availability of the references file from all registered repositories (all repositories corresponding to the repository-[repositoryId] files available) and aborts otherwise.
    • All the references available are collected.
    • All the blobs available in the DataStore are collected and deletion candidates identified by calculating all the blobs available not appearing in the blobs referenced. Only blobs older than a specified time interval from the earliest available references file are deleted. (last modified say 24 hrs (default)). The earliest references are identified by means of a timestamp marker file (markedTimestamp-[repositoryId]) for each repository.

The shared DataStore garbage collection is applicable for the following DataStore(s):

  • FileDataStore
  • SharedS3DataStore - Extends the S3DataStore to enable sharing of the data store with multiple repositories

Checking GC status for Shared DataStore Garbage Collection

The status of the GC operations on all the repositories connected to the DataStore can be checked by calling:

  • MarkSweepGarbageCollector#getStats() which returns a list of GarbageCollectionRepoStats objects having the following fields:
    • repositoryId - The repositoryId of the repository
      • local - repositoryId tagged with an asterix(*) indicates whether the repositoryId is of local instance where the operation ran.
    • startTime - Start time of the mark operation on the repository
    • endTime - End time of the mark operation on the repository
    • length - Size of the references file created
    • numLines - Number of references available
  • If the MBeans are registered in the MBeanServer then the following can also be used to retrieve the status:
    • BlobGC#getBlobGCStatus() which returns a CompositeData with the above fields.

This operation can also be used to ascertain when the ‘Mark’ phase has executed successfully on all the repositories, as part of the steps to automate GC in the Shared DataStore configuration. It should be a sufficient condition to check that the references file is available on all repositories. If the server running Oak has remote JMX connection enabled the following code example can be used to connect remotely and check if the mark phase has concluded on all repository instances.

import java.util.Hashtable;

import javax.management.openmbean.TabularData;
import javax.management.MBeanServerConnection;
import javax.management.MBeanServerInvocationHandler;
import javax.management.ObjectName;
import javax.management.remote.JMXConnectorFactory;
import javax.management.remote.JMXServiceURL;
import javax.management.openmbean.CompositeData;


/**
 * Checks the status of the mark operation on all instances sharing the DataStore.
 */
public class GetGCStats {

    public static void main(String[] args) throws Exception {
        String userid = "<user>";
        String password = "<password>";
        String serverUrl = "service:jmx:rmi:///jndi/rmi://<host:port>/jmxrmi";
        String OBJECT_NAME = "org.apache.jackrabbit.oak:name=Document node store blob garbage collection,type=BlobGarbageCollection";
        String[] buffer = new String[] {userid, password};
        Hashtable<String, String[]> attributes = new Hashtable<String, String[]>();
        attributes.put("jmx.remote.credentials", buffer);
        MBeanServerConnection server = JMXConnectorFactory
            .connect(new JMXServiceURL(serverUrl), attributes).getMBeanServerConnection();
        ObjectName name = new ObjectName(OBJECT_NAME);
        BlobGCMBean gcBean = MBeanServerInvocationHandler
            .newProxyInstance(server, name, BlobGCMBean.class, false);

        boolean markDone = checkMarkDone("GlobalMarkStats", gcBean.getGlobalMarkStats());
        System.out.println("Mark done on all instances - " + markDone);
    }

    public static boolean checkMarkDone(String operation, TabularData data) {
        System.out.println("-----Operation " + operation + "--------------");

        boolean markDoneOnOthers = true;
        try {
            System.out.println("Number of instances " + data.size());

            for (Object o : data.values()) {
                CompositeData row = (CompositeData) o;
                String repositoryId = row.get("repositoryId").toString();
                System.out.println("Repository  " + repositoryId);

                if ((!row.containsKey("markEndTime")
                        || row.get("markEndTime") == null
                        || row.get("markEndTime").toString().length() == 0)) {
                    markDoneOnOthers = false;
                    System.out.println("Mark not done on repository : " + repositoryId);
                }
            }
        } catch (Exception e) {
            System.out.println(
                "-----Error during operation " + operation + "--------------" + e.getMessage());
        }
        System.out.println("-----Completed " + operation + "--------------");

        return markDoneOnOthers;
    }
}
Unregistration

If a repository no longer shares the DataStore then it needs to be unregistered from the shared DataStore by following the steps:

  • Identify the repositoryId for the repository using the steps above.
  • Remove the corresponding registered repository file (repository-[repositoryId]) from the DataStore
    • FileDataStore - Remove the file from the data store root directory.
    • S3DataStore - Remove the file from the META folder of the S3 bucket.
    • AzureDataStore - Remove the file from the META folder of the Azure container.
  • Remove other files corresponding to the particular repositoryId e.g. markedTimestamp-[repositoryId] or references-[repositoryId].
Reset Repository ID

If any of the repositories sharing a DataStore is cloned then it would have the same repositoryId registered in the datastore as the repository from which cloned and running a shared DSGC would lead to data loss/missing blobs. So, care must be taken to reset the repositoryId.

oak-run cli utility has a resetclusterid command to reset the repository id. The command can be executed as below and more details can be found at the readme for oak-run:

java -jar oak-run.jar resetclusterid < repository path | Mongo URI >

Consistency Check

The data store consistency check will report any data store binaries that are missing but are still referenced. The consistency check can be triggered by:

  • MarkSweepGarbageCollector#checkConsistency
  • If the MBeans are registered in the MBeanServer then the following can also be used:
    • BlobGCMbean#checkConsistency

After the consistency check is complete, a message will show the number of binaries reported as missing. If the number is greater than 0, check the logs configured for org.apache.jackrabbit.oak.plugins.blob .MarkSweepGarbageCollector for more details on the missing binaries.

Below is an example of how the missing binaries are reported in the logs:

11:32:39.673 INFO [main] MarkSweepGarbageCollector.java:600 Consistency check found 1 missing blobs 11:32:39.673 WARN [main] MarkSweepGarbageCollector.java:602 Consistency check failure in the the blob store : DataStore backed BlobStore [org.apache.jackrabbit.oak.plugins.blob.datastore.OakFileDataStore], check missing candidates in file /tmp/gcworkdir-1467352959243/gccand-1467352959243

  • datastorecheck command of oak-run can also be used to execute a consistency check on the datastore. The details on how to execute the command and the different parameters can be checked in the readme for the oak-run module.