Vlad

Articles

YT Cluster

Every node accepts writes. On the Postgres you already run. YT Cluster: vanilla PostgreSQL plus an extension and a server process - no forked engine, no proxy in front, no managed-cloud lock-in - dbinvent.com

Released - tested and ready:

• No data transfer - your database becomes a cluster node where it already sits; nothing is dumped, copied or migrated
• No vendor lock - remove the extension and you are back to a single plain PostgreSQL, with your data in place
• No code change - same SQL, same drivers, same extensions; a plain COMMIT is a cluster commit, with no PgBouncer and no explicit completion call
• No dedicated master - nothing to promote, so nothing to fail over; a RAFT-based coordinator arbitrates and assigns workers, and never becomes the bottleneck
• Writes spread across every node - transaction boundaries are derived at runtime (virtual dynamic partitioning), so disjoint writes land on different nodes
• Synchronization as configuration - per-table rules decide what is redundant, what is local, and how many nodes a commit waits for: half, nodes>=N, or asynchronous
• Resiliency in three layers - two-phase commit on the way in; self-healing after the fact, which re-reads, votes on and repairs a divergent copy; and a row, table or node nobody can vouch for is blocked rather than guessed
• Software-defined topology (AZ, zones, tiers) - replication scope and row placement are rules, rewritten online with no redeploy or downtime
• MPP analytics - Apache DataFusion planned against the declared topology, ClickHouse for columnar workloads; neither needs a copy of the data
• Asking in plain language - yt_llm('...') on a local model returns PostgreSQL SQL, ClickHouse SQL or a DataFusion script, chosen by where the data lives; the model sees the catalog, never a row (how it works)
• One package, not a suite - server functions (RPPD), etcd key-value, streaming and queue (etcd) ship with the cluster, and it applies schema migrations with Schema Guard’s library - the sge tool itself is an open-source package of its own
• C and Rust - the extension in C, the server and tools in Rust

Measured, not asserted: 1.7x the throughput (171%) of PostgreSQL synchronous replication on a write-only load and 1.6x (162%) on an 80/20 read/write mix, at equal replication scope. Full synchronous two-phase commit across every node costs more - it is there for the tables that need it, not the default.

Articles

• From 30 years of personal IT experience into 30 pillars of the solution: 30 pillars
• What replication actually costs: 2PC
• Software-defined topology, verified: Where Rows Live
• Linear scalability possible: CAP
• DNS for service discovery: DNS
• Questions in plain language, answered by PostgreSQL, DataFusion or ClickHouse: LLM
• Resilience and membership under change, on three linked clusters: RMC (results)

Run it

Three writable nodes, on your laptop, in one command - the package comes from the public repository:

git clone https://github.com/DBinvent/yt-cluster-tests.git
cd yt-cluster-tests
APT_HOST=apt.dbinvent.com APT_HOST_IP= ./docker.sh

# then, inside the container - a three-node cluster:
cargo make --env PG_VER=14 --env CNT=3

Point psql at any node, write on two at once, and watch the rows appear on the third.

https://dbinvent.com/cluster/