Fork me on GitHub

Hybrid Index

New in 1.6

In Oak 1.6 (OAK-4412), we add support for near real time (NRT) and limited support for sync indexes. In OAK-6535, we add support for sync property indexes. See also Oakathon August 2017 PresentationHybrid Index v2.pdf

Synchronous Index Usecases

Synchronous indexes are required in the following use-cases:

Unique Indexes

For unique indexes like the uuid index, and the principal name index, we need to be ensured that an indexed value is unique across the whole repository on commit. If the indexed value already exists, e.g. principal with same name already exist, then the commit should fail. For this, we need a synchronous index, which get updated as part of commit itself.

Property Indexes

Depending on application requirements the query results may be:

  • Eventually Consistent - Any changes eventually gets reflected in query results.
  • Consistent - Any change immediately gets reflected in query results.

For most cases, for example user-driven search, eventual consistent search result work fine, and hence async indexes can be used. With NRT indexes (OAK-4412), changes done by user get reflected very soon in search results.

However, for some cases we need to support fully consistent search results. Assume there is component which maintains a cache for nodes of type app:Component, and uses a observation listener to listen for changes in nodes of type app:Component, and upon finding any changes, it rebuilds the cache by queriying for all such nodes. For this cache to be correct, it needs to be ensured query results are consistent with the session associated with the listener. Otherwise it may miss a new component, and a later request to the cache for such component would fail.

For such use-cases, it's required that indexes are synchronous and results provided by index are consistent.

Drawbacks of current property indexes

Oak has support for synchronous property indexes, which are used to meet above use-cases. However the current implementation has certain drawbacks:

  • Slow reads over remote storage - The property indexes are stores as normal NodeState and hence reading them over remote storage like Mongo performs poorly (with Prefetch, OAK-9780, this is improved).
  • Storage overhead - The final storage overhead is larger, specially for remote storage where each NodeState is mapped to 1 document. (On the other hand, temporary disk usage for Lucene indexes might be higher than for node stores, due to write amplification with Lucene.)

Proposal

To overcome the drawbacks and still meet the synchronous index requirements, we can implement a hybrid index where the indexes content is stored using both property index (for recent entries) and Lucene indexes (for older entries):

  • Store recently added index content as a normal property index.
  • As part of async indexer, store the content in the Lucene index.
  • Later prune the property index content (parts that have been indexed in Lucene).
  • Any query is a union of query results from both property index and Lucene indexes (with some caveats).

Index Definition

The synchronous index support needs to be enabled via index definition:

  • Set sync to true for each property definition which needs to be indexed in a sync way
/oak:index/assetType
- jcr:primaryType = "oak:QueryIndexDefinition"
- type = "lucene"
- async = ["async"]
+ indexRules
    + nt:base
        + properties
            + resourceType
                - propertyIndex = true
                - name = "assetType"
                - sync = true

For unique indexes set unique i.e. true:

/oak:index/uuid
- jcr:primaryType = "oak:QueryIndexDefinition"
- type = "lucene"
- async = ["async"]
+ indexRules
    + nt:base
        + properties
            + uuid
                - propertyIndex = true
                - name = "jcr:uuid"
                - unique = true

Index Storage

The property index content is stored as hidden nodes under the index definition nodes. The storage structure is similar to property indexes with some changes.

Unique Indexes

/oak:index/assetType
+ :data   //Stores the lucene index files
+ :property-index
    + uuid
        + <value 1>
            - entry = [/indexed-content-path]
            - jcr:created = 1502274302 //creation time in millis
        + 49652b7e-becd-4534-b104-f867d14c7b6c
            - entry = [/jcr:system/jcr:versionStorage/63/36/f8/6336f8f5-f155-4cbc-89a4-a87e2f798260/jcr:rootVersion]

Here:

  • :property-index - hidden node under which property indexes is stored for various properties which are marked as sync.
  • For unique indexes, each entry also has a timestamp which is later used for pruning.

Property Indexes

/oak:index/assetType
+ :data   //Stores the lucene index files
+ :property-index
    + resourceType
        - head = 2
        - previous = 1
        + 1
            - jcr:created = 1502274302 //creation time in millis
            - lastUpdated = 1502284302
            + type1
                + libs
                    + login
                        + core
                            - match = true
            + <value>
              + <mirror of indexed path>
        + 2
            - jcr:created = 1502454302
            + type1
                + ...

Here we create new buckets of index values which simplifies pruning. New buckets get created after each successful async indexer run, and older buckets get removed. The bucket have a structure similar to tje content mirror store strategy.

For each indexed property, we keep a head property which refers to the current active bucket. This is changed by IndexPruner. In addition, there is a previous bucket to refer to the last active bucket.

On each run of IndexPruner:

  • Check if IndexStatsMBean#LastIndexedTime is changed from last known time.
  • If changed then:
    • Create a new bucket by incrementing the current head value.
    • Set previous to current head.
    • Set head to new head value.
    • Set lastUpdated on previous bucket to now.
  • Remove all other buckets.

We require both head and previous bucket as there is some overlap between content in previous.

Index Pruner

Index Pruner is a periodic task prunes the index content. It uses the IndexStatsMBean#LastIndexedTime to determine upto which time the async indexer has indexed the repository, and then removes entries from the property index which are older than that time.

  • Property index - here pruning is done by creating a new bucket and then removing the older bucket.
  • Unique index - Here prunining is done by iterating over current indexed content and removing the older ones.

Query Evaluation

On the query side, we perform a union query over the 2 index types: A union cursor is created which consist of:

  • LucenePathCursor - Primary cursor backed by Lucene index.
  • PropertyIndexCursor - A union of path cursor from current head and previous bucket.

Open Points

If there are multiple property definition in a Lucene index marked with sync and a query involves constraints on more than 1, then which property index should be picked is not clear.

Attachments:

Hybrid Index v2.pdf (application/pdf)