BookKeeper Table Service

BookKeeper Table Service is a contrib module added to BookKeeper, providing a table (aka key/value) service as part of the stream storage for bookkeeper.

Detail Design

BP-30

Build

$ mvn clean install -DskipTests

Start Standalone

$ bin/streamstorage standalone

CLI Usage

CLI is available at bin/streamstorage-cli

Examples

Distributed Counters

Create Table

Stream is the management unit. A Table is a materialized view of a Stream.

bin/streamstorage-cli -s 127.0.0.1:4181 stream create --stream test_dist_counter

Listen on a key

bin/streamstorage-cli -s 127.0.0.1:4181 table get -t test_dist_counter -k "counter-1" --watch

Increment a counter

bin/streamstorage-cli -s 127.0.0.1:4181 table incr -t test_dist_counter -k "counter-1" -a 100

K/V Store

Use the table service as the normal k/v store for storing metadata

Create Table

bin/streamstorage-cli -s 127.0.0.1:4181 stream create --stream test_kv_store

Listen on a key

bin/streamstorage-cli -s 127.0.0.1:4181 table get -t test_kv_store -k "test-key" --watch

Put Key/Value

bin/streamstorage-cli -s 127.0.0.1:4181 table put -t test_kv_store -k "test-key" -v "test-value-`date`"

Increment

bin/streamstorage-cli -s 127.0.0.1:4181 table get -t test_kv_store -k "test-counter-key" --watch

bin/streamstorage-cli -s 127.0.0.1:4181 table incr -t test_kv_store -k "test-counter-key" -a 200

Features

• API Model: support PTable & Table
  • PTable: short for partitioned table. the table is modeled as <pKey, lKey> -> value, kv pairs are partitioned based on pKey. range operations over a single pKey is supported.
    • put/get/delete on single <pKey, lKey>
    • range/deleteRange on a single <pKey>
    • txn on a single <pKey>
    • increment on single <pKey, lKey>
  • Table: the table is modeled as <key> -> value. kv pairs are also partitioned, based on key. range operations are not supported. single key txn (aka cas operation) is supported.
• Persistence
  • The source-of-truth of a table is its journals. The journal is a stream comprised of multiple log segments (aka ledgers).
  • Rocksdb as its materialized index for each range partition. Rocksdb can be ephemeral and it can be restored from checkpoints that are also persisted as log streams.
  • Rocksdb can spill in-memory data to disk. Additionally, the rocksdb files are periodically incrementally checkpointed to bookkeeper for fast recovery.
• Clients
  • gRPC & protobuf based
  • Thick client with client-side request routing
  • Java implementation
  • Thin client with server-side request routing
  • Multiple language clients: C++, Python, Go
• Deployment
  • Run table service as a bookkeeper lifecycle component in bookie server

Later Improvements

• Stream <-> Table exchangeable: (for metadata store)
  • Can retrieve a Stream of updates from a key, a range of keys or a Table.
  • A Stream can be materialized and viewed as a Table.
• TTL or Lease: for supporting membership (for metadata store)
• Auto Scale: split and merge ranges. should apply for both Stream and Table.