Making DocDB High-Performance

DocDB uses a highly customized version of RocksDB, a log-structured merge tree (LSM) based key-value store.

A number of enhancements or customizations were done to RocksDB in order to achieve scale and performance. These are described below.

Enhancements to RocksDB

Efficiently model documents

To implement a flexible datamodel—-such as a row (comprising of columns), or collections types (such as list, map, set) witharbitrary nesting—-on top of a key-value store, but, more importantly, to implement efficientoperations on this data model such as:

• fine-grained updates to a part of the row or collection without incurring a read-modify-writepenalty of the entire row or collection
• deleting/overwriting a row or collection/object at an arbitrary nesting level without incurring aread penalty to determine what specific set of KVs need to be deleted
• enforcing row/object level TTL based expiry:

A tighter coupling into the “read/compaction” layers of the underlying KV store (RocksDB) is needed.We use RocksDB as an append-only store and operations such as row or collection delete are modeledas an insert of a special “delete marker”. This allows deleting an entire subdocument efficientlyby just adding one key/value pair to RocksDB. Read hooks automatically recognize these markers andsuppress expired data. Expired values within the subdocument are cleaned up/garbage collected by ourcustomized compaction hooks.

Raft vs RocksDB WAL logs

DocDB uses Raft for replication. Changes to the distributed system are already recorded or journaled as part of Raft logs. When a change is accepted by a majority of peers, it is applied to each tablet peer’s DocDB, but the additional WAL mechanism in RocksDB (under DocDB) is unnecessary and adds overhead.For correctness, in addition to disabling the WAL mechanism in RocksDB, YugabyteDB tracks the Raft“sequence id” up to which data has been flushed from RocksDB’s memtables to SSTable files. Thisensures that we can correctly garbage collect the Raft WAL logs as well as replay the minimal numberof records from Raft WAL logs on a server crash or restart.

MVCC at a higher layer

Multi-version concurrency control (MVCC) in DocDB is done at a higher layer, and does not use the MVCC mechanism of RocksDB.

The mutations to records in thesystem are versioned using hybrid-timestamps maintained at the YBase layer. As a result, the notionof MVCC as implemented in a vanilla RocksDB (using sequence ids) is not necessary and only addsoverhead. YugabyteDB does not use RocksDB’s sequence ids, and instead uses hybrid-timestamps that arepart of the encoded key to implement MVCC.

Backups and snapshots

These need to be higher level operations that take into consideration data inDocDB as well as in the Raft logs to get a consistent cut of the state of the system

Data model aware Bloom filters

The keys stored by DocDB in RocksDB consist of a number of components, where the first component is a “document key”, followed by a few scalar components, and finally followed by a timestamp (sorted in reverse order).

The bloom filter needs to be aware of what components of the key need be added to the bloom so that only the relevant SSTable files in the LSM store are looked up during a read operation.

In a traditional KV store, range scans do not make use of bloom filters because exact keys that fall in the range are unknown. However, we have implemented a data-model aware bloom filter, where range scans within keys that share the same hash component can also benefit from bloom filters. For example, a scan to get all the columns within row or all the elements of a collection can also benefit from bloom filters.

Range query optimizations

The ordered (or range) components of the compound-keys in DocDB frequently have a natural order. For example, it may be an int that represents a message id (for a messaging like application) or a timestamp (for a IoT/Timeseries like use case). See example below. By keeping hints with each SSTable file in the LSM store about the min/max values for these components of the “key”, range queries can intelligently prune away the lookup of irrelevant SSTable files during the read operation.

SELECT message_txt
  FROM messages
WHERE user_id = 17
  AND message_id > 50
  AND message_id < 100;

Or, in the context of a time-series application:

SELECT metric_value
  FROM metrics
WHERE metric_name = ’system.cpu’
  AND metric_timestamp < ?
  AND metric_timestamp > ?

Efficient memory usage

There are two instances where memory usage across components in a manner that is global to the server yields benefits, as described below.

Server-global block cache

A shared block cache is used across theDocDB/RocksDB instances of all the tablets hosted by a YB-TServer. This maximizes the use of memoryresources, and avoids creating silos of cache that each need to be sized accurately for differentuser tables.

Server-global memstore limits

While per-memstore flush sizes can beconfigured, in practice, because the number of memstores may change over time as users create newtables, or tablets of a table move between servers, we have enhanced the storage engine to enforce aglobal memstore threshold. When such a threshold is reached, selection of which memstore to flushtakes into account what memstores carry the oldest records (determined using hybrid timestamps) andtherefore are holding up Raft logs and preventing them from being garbage collected.

Scan-resistant block cache

We have enhanced RocksDB’s block cache to be scan resistant. The motivation was to prevent operations such as long-running scans (e.g., due to an occasional large query or background Spark jobs) from polluting the entire cache with poor quality data and wiping out useful/hot data.