Lore

A timeline of what shaped the CockroachDB codebase. Dates are derived from git history (first commit, tag dates, and earliest commits inside relevant directories). Where a "why" is uncertain it is hedged accordingly.

Eras

The Spencer Kimball era — Feb 2014 to ~2015

The first commit (2014-02-05, by Spencer Kimball) introduced a project named after a creature that "scales horizontally and survives". The original design doc (docs/design.md) is still in-tree. The very first cut was already aiming at a Spanner-style architecture: a single sorted KV map, ranges replicated by Raft, snapshot isolation by default. Most of the foundational packages (pkg/kv, pkg/storage, pkg/util, pkg/gossip) date from this period.

About 1,344 commits landed in 2014 and another 5,848 in 2015 as the team grew toward beta.

The first-stable rush — 2016 to mid-2017

By 2016 the project had a SQL layer, distributed transactions, and the original RocksDB-backed engine. The v1.0 release was tagged in 2017. Activity climbed to ~9,000–10,000 commits/year and the team expanded substantially.

Distributed SQL and the optimizer — 2017 to 2019

pkg/sql/distsql* and pkg/sql/flowinfra/ show how SQL execution went from a single-node row engine to a distributed flow of processors. The cost-based optimizer in pkg/sql/opt/ was built on top, replacing the original heuristic planner for most query shapes. The columnar/vectorized engine (pkg/sql/colexec/, pkg/sql/colflow/) followed shortly after as a way to compete on analytical workloads.

The Pebble migration — ~2019 to 2020

CockroachDB replaced its RocksDB engine with Pebble, a native Go LSM-tree maintained by the storage team. The transition is visible across pkg/storage/pebble*.go and pkg/storage/min_version.go, which now ratchets Pebble's "format major version" alongside the cluster version.

The multi-region and serverless build-out — 2020 to 2022

Three large initiatives reshaped large parts of the tree:

• Multi-region SQL (pkg/ccl/multiregionccl/, pkg/sql/catalog/multiregion/, pkg/sql/alter_table_locality.go) — REGIONAL BY ROW, ZONE, GLOBAL tables, regional survivability.
• Multi-tenancy (pkg/multitenant/, pkg/ccl/multitenantccl/, pkg/server/server_controller*.go) — many SQL servers can share a host KV layer, eventually enabling CockroachDB Serverless.
• Span configurations (pkg/spanconfig/) — zone configs were rewritten as a span-keyed reconciliation system that scales with the table count.

The schema changer rewrite — 2022 to today

A "declarative" schema changer (pkg/sql/schemachanger/) is gradually replacing the legacy schema-changer. As of pkg/sql/794d7ae (Apr 2026), an in-tree document explicitly contrasts the legacy and type-schema-changer architectures while both still co-exist. Both implementations live in-tree.

Store-liveness leases and disaggregated storage — 2023 to today

Two long-running modernizations dominate recent commits:

• Store-liveness leases (pkg/kv/kvserver/storeliveness/, pkg/raft/raftstoreliveness/) — a new leadership/lease primitive that decouples leader elections from node liveness gossip, replacing epoch-based leases.
• Disaggregated storage (pkg/storage/shared_storage.go, pkg/storage/external_sst_reader.go, pkg/cloud/) — pushing Pebble-managed L5/L6 SSTs into shared object storage, primarily to reduce inter-AZ replication cost.

Cross-cluster replication — 2023 to today

The pkg/crosscluster/ and pkg/ccl/streamingccl/-family code grew into a complete logical and physical replication offering: PCR (physical cluster replication) and LDR (logical data replication) both run as long-lived jobs that read rangefeeds out of one cluster and ingest them into another.

Longest-standing code

• pkg/kv/doc.go, pkg/storage/doc.go — both carry 2014 copyright headers and describe APIs that, while heavily evolved, still exist.
• pkg/util/hlc/ — Hybrid Logical Clocks were one of the earliest contributions and remain the cluster's time primitive.
• pkg/gossip/ — the gossip network has survived essentially every era; its shape is mostly unchanged.
• pkg/raft/ (formerly imported from etcd) — first imported in 2015, then forked in-tree to support epoch leases, store liveness, and per-replica term caches. The fork still includes the original LICENSE and README.md.

Deprecated and removed features

• RocksDB engine — replaced by Pebble. References in old comments still mention RocksDB; the engine code is gone.
• PROTECTED_TIMESTAMP v1 — superseded by the pkg/protectedts/-style implementation; old job records may still appear in mixed-version clusters.
• Epoch-based leases are being phased out in favor of store-liveness leases (pkg/kv/kvserver/replica_store_liveness.go).
• Old jobs adoption protocol — comments in pkg/jobs/registry.go and pkg/jobs/adopt.go reference an older claim/lease model that has been rebuilt around sqlliveness sessions.
• The legacy schema changer lives next to its replacement; tables that touch many subsystems still occasionally fall back to it.

Major rewrites

What	When	What replaced it
RocksDB engine	~2019–2020	Pebble in pkg/storage/pebble.go
Heuristic planner	2017–2019	Cost-based optimizer in pkg/sql/opt/
Row-only execution	2018–2020	Vectorized engine in pkg/sql/colexec/, pkg/sql/colflow/
Zone-config dispatch	2021–2022	pkg/spanconfig/ reconciler
Legacy schema changer	2022–today	pkg/sql/schemachanger/ (declarative)
Epoch leases	2023–today	pkg/kv/kvserver/leases/ + store-liveness
Single-tenant server	2021–today	Multi-tenant pkg/server/server_controller*.go

Growth trajectory

Commits per year on master:

2014:  1,344    2018: 10,477    2022: 11,480
2015:  5,848    2019:  8,020    2023: 12,488
2016:  9,697    2020: 10,685    2024: 11,380
2017:  9,142    2021: 10,958    2025: 11,254
                                2026:  4,910 (partial)

The codebase has roughly tripled in size since 2017, but the commit cadence has been remarkably steady at ~10–12k commits/year. pkg/cmd/ and pkg/ui/ grew along with the operational and observability story; pkg/ccl/ grew with every CSL feature added.