databases

databases you might consider using

databases

People keep asking me career advice, here is one:

Make state someone else's problem.

— Jaana Dogan (@rakyll) 2020-05-19

jepsen.io analyses arre imformative

basics

common-ish databases you might think about using

slight focus on distributed databases

isolation and consistency

Read isolation and consistency, and this and problems that may occur (the last table is a simple overview) tldr: it's confusing, more so when transactions are involved

Simple way to think about it, add external consistency and stric serializable above serializable in isolation

• external consistency: behave as if a single machine, regardless of replica
• strict serializable: add realtime ordering constraint on serial order
• serializable: behave as if transactions occurred in some serial order
• snapshot (repeatable read): view a snapshot taken when transaction begins
• read committed: view things after they have committed,
• read uncommitted: view as things change, even if they will get rolled back

CAP

Choose 2, only google can have CA

• Consistency
• Availability
• Partitions

relational

The standard database.

embedded relational

not much choice here

notes

• Sqlite needs WAL (write ahead log) for snapshot
• Sqlite has no isolation in a single connection
• modernc.org/sqlite is a translation of C to Go

single node relational

postgres is usually recommended

notes

• PostgreSQL License is similar to MIT / BSD
• some of the multi node ones below can be run in single node mode too

multi node relational

notes

• CockroachDB and yugabyte db both claim to be modelled after spanner
• CockroachDB Business Source License converts to Apache 2.0 3 years after release
• dqlite is built on sqlite
• Galera is the clustering mode for mysql / mariadb
• Galera is serializable in master-slave, snapshot in multi master
• rqlite is built on sqlite
• Spanner is technically CP
• Vitess is designed for sharded data
• YugaByte universe = primary cluster + read clusters
• YugaByte cluster = multi node auto sharded
• YugaByte is serializable in YSQL (sql), snapshot in YCQL (nosql)

key value

Stores a blob of unindexed data.

embedded key value

notes

• badger does not resolve conflicts, user retries
• bbolt is single process
• LevelDB isolation is probably serializable since only a single writer is allowed
• goleveldb appears incomplete
• LMDB OpenLDAP Public License is similar to MIT / BSD
• RocksDB actually allows concurrent processes
• pebble allows multi process

single node key value

notes

• Redis is primarily an in memory cache, durability is not guaranteed
• Redis core is BSD 3, redis modules are Redis Source Available License
• Redis technically has cluster / replication but it's async (useful for sometimes stale caching)
• Redis isolation is probably serializable

multi node key value

notes

• Consul external consistency requires extra roundtrips
• etcd external consistency requires extra roundtrips

document

Stores structured data (can be indexed). Can be used like a key-value store.

notes

• ArangoDB is multimodal but nobody recommends the other modes
• CouchDB is eventually consistent
• MongoDB loses data by default
• RethinkDB is probably serializable

wide column

Stores column data contiguously (as opposed to row data contiguously). Primary usecase is analytics.

notes

• BigTable is single node
• Google internally has Percolator with snapshot isolation built on BigTable
• Cassandra can optionally be CP
• Cassandra only has row level isolation
• HBase requires hadoop

graph

Stores nodes and edges, super specialized.

notes

• Neo4j cluster also has primary-read only

time series

Write once, append only style. No transactions, no isolation concerns.

notes

• Druid is a column database but doesn't support writes
• Graphite has a TSDB but cannot be operated independently, also not clustered
• InfluxDB clustering is available as enterprise offering
• OpenTSDB requires HBase, Hadoop
• TimescaleDB is based on postgres
• TimescaleDB clustering is private beta