Status

Current state: [ UNDER DISCUSSION ]

Discussion thread: <link to mailing list DISCUSS thread>

JIRA: SAMZA-1278

Released: 

Problem

Samza today supports RocksDB and MemDB as local data stores, which enables users to cache adjunct data for later usage during stream processing. However, the population of a data store is end user’s responsibility. This introduced additional effort from end user to develop and maintain data stores, and to deal with corner cases such as reload after consumers falling off. To avoid these issues, some people employed alternative solutions such as Voldemort, CouchBase, etc. In addition, table oriented operations of fluent API would require working data to be made available by the system. As we look at the issue more closely, it appears generic enough to be addressed by data infrastructure.

A dataset delivered in a stream can be either bounded or unbounded, an example of an unbounded dataset could be a database change stream, and an example of a bounded dataset could be the content of a file. When Samza is running in 24x7 mode, the stream for a bounded dataset may deliver multiple versions.

This proposal focuses on unbounded datasets.

Motivation

We want to have an adjunct data (AD) store that is a read-only cache. It automatically stores streaming data for later usage. Adjunct data can be accessed the same way as accessing a key-value store in Samza, we provide a consistent view of data from a Samza task’s or container's perspective. Data can be either partitioned or unpartitioned. If the dataset is small enough to fit in a RocksDB instance, the same copy would be populated in every container via a broadcast stream; if it is too large fit in one database instance it would be partitioned across containers of a Samza job. 

Proposed Changes

The proposed changes include

• Introduce a new construct AdjunctDataStoreManager, its responsibility includes
  • Maintains mapping between system streams and adjunct data stores
  • Extract the key and value from an IncomingMessageEnvelop
  • Save data in stores
• TaskInstance will be modified to consult AdjunctDataStoreManager before delivering a message to a task; a message is stored in an adjunct store, it is not delivered to tasks
• Introduce new configurations to "associate" a Samza K/V store to a system stream
• Provide hooks to transform an incoming message to desired types (this is useful to store a subset of the incoming message)

Data partitioning

For small datasets that can easily fit in a local store, the AD stream can be configured as broadcast stream and delivered to all containers. Within a container, one AD store instance per stream is instantiated and shared among all tasks within that container. 

 For large datasets too large to fit in a local store, the AD stream has to be partitioned, there will be one AD store instance per stream and task. 

Consistency

For unbounded dataset, the version of the dataset is naturally consistent across containers (as there is only one). If an stream is unpartitioned, we provide a consistent snapshot within a container. If a stream is partitioned, we provide a consistent snapshot of a version within a task. No consistency is offered across containers.

Bootstrap

When an AD stream is marked as bootstrap, it guarantees that an initial snapshot is built before processing of input streams starts; otherwise input stream and AD streams are processed at the same time. After bootstrap, for change capture data, we keep updating its AD store when new updates arrives.

Key and Value

By Default, we use the key and value in IncomingMessageEnvelop to store an AD data item in a store, user provides serde as done today. Configuration options will be provided to allow user to

• construct user defined keys
• construct user defined values

User defined objects will be treated as POJO's, and we may provide a default serde for POJO.

Configuration

Key classes

AdjunctDataStoreManagerFactory

/**
 * Factory class for AdjunctDataStoreManager
 */
public interface AdjunctDataStoreManagerFactory {
    AdjunctDataStoreManager getAdjunctDataStoreManager(
        Config config, 
        Set<SystemStreamPartition> ssps, 
        Map<String, StorageEngine> storageEngines);
}

AdjunctDataStoreManager

/**
 * An AdjunctDataStoreManager instance is responsible for
 * 1. maintaining the mapping between system streams and adjunct data stores
 * 2. extracting the key and value from an IncomingMessageEnvelop
 * 3. populating adjunct data stores
 */
public interface AdjunctDataStoreManager {
   /**
    * Invoked before a message is passed to a task
    * @returns true if the message is saved in a store, false otherwise
    */
    AdjunctDataStoreManager boolean process(IncomingMessageEnvelop message);
}

/**
 * A factory to instantiate key/value extractors
 */
public interface KeyValueExtractorFactory {
    KeyValueExtractor getKeyValueExtractor(Config config);
}

/**
 * A key/value extractor that extracts key and value from in incoming
 * message envelop
 */
public interface KeyValueExtractor {
    Object getKey(IncomingMessageEnvelop input);
    Object getValue(IncomingMessageEnvelop input);
}

No changes to public interface

Implementation and Test Plan

• Introduce interfaces and implementation described above
• Add unit tests to test and verify functionality (plain, bootstrap, broadcast, ...)
• Verify in different deployment environment, note: implementation should be agnostic to deployment environment (YARN, Hadoop or standalone). 

Compatibility, Deprecation, and Migration Plan

As this is a new feature, no plans are required for compatibility, deprecation and migration.

Rejected Alternatives

1. One AD store instance per stream - this option guarantees a consistent view within a container with side effects. Since data in AD streams may arrive in different times, the maintenance of a consistent snapshot becomes difficult. As multiple versions of a dataset evolves, we potentially may have to maintain a large number of intermediate snapshot versions.

2. One AD store instance per stream partition - this option guarantees a consistent view at stream partition level. Querying the store becomes a problem: when multiple partitions are assigned to a task, we have no way of knowing which store instance holds the value resides and end up querying all instances. The implementation might be a bit simpler, but the price is very high.

3. One AD store instance per stream and task - this is the compromise between 1 and 2, here we guarantee a consistent view per task. If multiple partitions are assigned to a task, we potentially still have to maintain multiple intermediate snapshot versions of a AD store instance, but number of versions is lower than in #1. Since within LI no one is using the custom system partition grouper feature, only open source portion will be impacted.

#3 chosen

AD store abstraction

For bounded datasets delivered (such as files) as streams, the adjunct data store may need to maintain multiple version of underlying stores in order to guarantee serving of a consistent version. This would require the abstraction of adjunct data store, which manages internally versions of underlying stores.

Key extractor and value converter

Another option considered is to embed key extraction and value conversion logic in serde layer. It would work, but would make serde asymmetric and less portable.