I'm reimplementing Oracle RAC on PostgreSQL — in the open.

PostgreSQL has never had a shared-disk, multi-active cluster (its HA is shared-nothing replication). pgrac brings the Oracle RAC model — many nodes, one shared database, Cache Fusion / SCN / GES — to PostgreSQL 16.13.

⚠️ Early stage, built in public.

Per-feature status — working / scaffolded / planned — is tracked honestly at pgrac.dev/features. The short version:

Running today: the cluster substrate (TCP interconnect + 1 Hz LMON heartbeat, per-peer state in pg_cluster_ic_peers; SCN / ITL / dedicated-undo block format; multi-node pgrac.conf bootstrap), plus working code paths for Cache Fusion's 3-way block transfer, cross-node MVCC, the global SCN clock, cluster catalog invalidation, and a cluster-aware storage manager. Honest caveat: these run on real paths, but cross-node behavioral test coverage is still being built.

In progress / planned: full cross-node GES enqueue locking and crash recovery (active); RDMA transport provider/mux work, Active Data Guard, rolling upgrade, FAN/TAF, and backup/DR.

Sanity anchor: the --disable-cluster build is binary-identical to upstream PostgreSQL 16.13 and passes the full 219-test regression suite.

⭐ If "Postgres with RAC-style shared storage" is something you've wanted, star the repo to follow the build.

Two PostgreSQL nodes, one cluster, a live TCP heartbeat interconnect — the per-peer heartbeat counters climb in real time:

Reproduce it locally in about a minute — see Multi-node cluster in the bootstrap guide.

Project site: pgrac.dev — architecture deep-dives, the full feature catalog, and a side-by-side comparison with Oracle RAC (coming online).

pgrac targets the full Oracle RAC model. Part of it runs today; much is still being built (see the status above) — these diagrams show the design.

Cluster topology — shared storage, interconnect, Cache Fusion, Active DG standby:

The stack — which layers are stock PostgreSQL vs. pgrac additions:

Cache Fusion — the 3-way GRD protocol (scaffolded today):

Cluster MVCC — global-SCN visibility + per-node undo (design):

More diagrams and deep-dives at pgrac.dev.

User-facing manual:

PostgreSQL upstream documentation lives under doc/ and is shipped unchanged from the upstream tree.

git clone https://github.com/sqlrush/pgrac.git
cd pgrac

./configure --prefix=$HOME/pgrac-install \
            --enable-cluster --enable-tap-tests \
            --with-openssl --with-icu --with-lz4 --with-zstd
make -j$(getconf _NPROCESSORS_ONLN 2>/dev/null || sysctl -n hw.ncpu)
make install

export PATH=$HOME/pgrac-install/bin:$PATH

pgrac-init -D /tmp/pgrac-demo --node-id=0 --cluster-name=demo
echo "port = 65433"                                  >> /tmp/pgrac-demo/postgresql.conf
echo "unix_socket_directories = '/tmp'"              >> /tmp/pgrac-demo/postgresql.conf
echo "listen_addresses = ''"                         >> /tmp/pgrac-demo/postgresql.conf

pgrac-start -D /tmp/pgrac-demo -l /tmp/pgrac-demo.log -w
psql -h /tmp -p 65433 -d postgres -c 'SELECT * FROM pg_cluster_nodes;'

See docs/user-guide/bootstrap.md for details.

The build follows the standard PostgreSQL configure + make + make install flow. Two extra flags are pgrac-specific:

• --enable-cluster activates the cluster subsystem.
• --enable-tap-tests enables the TAP test suites (Perl).

See docs/user-guide/install.md for the complete dependency list and step-by-step instructions on macOS and Linux.

pgrac is built in public and early — testing, feedback, and patches all help. See CONTRIBUTING.md and the good first issue label.

pgrac's cluster code is original work by WangYingJie sqlrush@gmail.com, licensed under the PostgreSQL License (BSD-style) — the same terms as PostgreSQL itself. See LICENSE and COPYRIGHT.

File issues at https://github.com/sqlrush/pgrac/issues.

Forked from PostgreSQL 16.13 (https://www.postgresql.org).