Cloud Parts

Alibaba Cloud PolarDB

Alibaba Cloud’s cloud-native relational databases: MySQL- and PostgreSQL-compatible clusters whose compute nodes share one copy of the data on distributed storage, plus PolarDB-X, a distributed MySQL-compatible edition.

What it is
Relational databases
Type
Commercial product
Made by
Alibaba Cloud
License
Proprietary managed service; open-source PolarDB for PostgreSQL (Apache 2.0) and PolarDB-X (Apache 2.0, storage engine GPL 2.0)
Facts checked
11 October 2026
Website Source code

What it is

A PolarDB cluster separates compute from storage. One primary node takes writes and up to 15 read-only nodes take reads, and none of them keeps its own copy of the data: data files and redo logs live on a shared, distributed storage layer (PolarStore) that keeps several replicas in step with Parallel-Raft, reached over RDMA. Adding a reader therefore adds only compute, with no data to copy. If the primary fails, a reader takes over in under 10 seconds with no data loss. A proxy in front (PolarProxy) splits reads from writes, pools connections and offers read consistency levels. PolarDB for MySQL runs engine versions 5.6, 5.7 and 8.0; the Multi-master Cluster (Limitless) Edition lets every node write, an in-memory column index (IMCI) serves analytics on the same data, serverless clusters scale in PolarDB Capacity Units (about 1 core and 2 GB of memory each), and the Global Database Network replicates across regions with under 2 seconds of lag. Storage grows automatically, to 500 TB per cluster by default.

PolarDB for PostgreSQL has the same shape and is compatible with PostgreSQL 11, 14, 15, 16 and 17. It replicates to its readers from redo logs, so even a long DDL doesn’t make them lag, a write on the primary is immediately visible to all readers, and the cluster endpoint uses each node’s log sequence number (LSN) so a read never lands on a node that hasn’t caught up. A sibling edition, PolarDB for PostgreSQL (Compatible with Oracle), accepts Oracle stored procedures, data types and built-in functions, and a distributed form splits data across compute and data nodes. PolarDB for Xscale, or PolarDB-X, is different: a shared-nothing, MySQL-compatible database whose stateless compute nodes route queries and coordinate two-phase commits, while data nodes replicate with Paxos and a metadata service hands out global timestamps (TSO) for distributed MVCC. It shards on the primary key by default, adds clustered columnar indexes for analytics, and can start as a single-node Standard Edition and be upgraded in place to the distributed Enterprise Edition.

The managed services run on Alibaba Cloud, with Apsara Stack and DBStack for private deployments. Two parts are open source on GitHub: PolarDB for PostgreSQL under Apache 2.0, with stable branches built on PostgreSQL 11, 15 and 17 (17.11.1.0 in August 2026), and PolarDB-X, whose compute layer is Apache 2.0 and whose MySQL 8.0-based storage engine is GPL 2.0, with a Kubernetes operator. In US (Virginia) a serverless PCU costs $0.0794 an hour on Enterprise Edition and $0.07 on Standard Edition. PSL4 storage outside mainland China costs $0.000542 per GB-hour pay-as-you-go (about $0.40 per GB-month) or $0.2569 per GB-month on subscription, with storage replicated across two zones; single-zone storage costs half as much.

Why teams pick it

  • Read capacity without copies: up to 15 readers share one copy of the data, adding one takes minutes and costs only compute, and LSN-based consistency keeps reads from going stale
  • Existing drivers and tools work: MySQL 5.6, 5.7 and 8.0, PostgreSQL 11 to 17, an Oracle-compatible edition, and MySQL-compatible PolarDB-X
  • Failover in under 10 seconds with no data loss, and backups of terabytes in minutes from storage snapshots, without locking tables
  • Several ways to grow past one primary: serverless scaling, multi-writer Limitless clusters, PolarDB-X sharding, and column indexes for analytics on live data
  • The PostgreSQL edition and PolarDB-X are open source, so you can try them in Docker or run them on Kubernetes outside Alibaba Cloud

Watch out for

  • The managed service is Alibaba Cloud’s (or its Apsara Stack and DBStack private clouds), and PolarDB for MySQL has no open-source release in the polardb GitHub organization, which carries only the PostgreSQL edition and PolarDB-X
  • PolarDB for MySQL offers engine versions 5.6, 5.7 and 8.0 (docs updated June 2026), while MySQL’s current long-term support series are 8.4 and 9.7
  • A Cluster Edition has one writer; scaling writes means the Limitless Edition or PolarDB-X, whose distributed Enterprise Edition is “highly compatible” with MySQL rather than fully compatible like its Standard Edition
  • Many price dimensions: edition, compute specification or serverless PCUs, storage class (ESSD PL0 to PL3, AutoPL, PSL4, PSL5), zone layout and region, and pay-as-you-go PSL4 storage costs about 1.5 times the subscription rate
  • Headline figures such as up to 6 times the transaction throughput and 400 times the analytical performance of open-source databases are Alibaba’s own

Alternatives

What all relational databases have in common →
CedarDBCommercialA PostgreSQL-compatible database rebuilt from scratch at TU Munich, running transactions and analytics on one copy of the data on a single server.Cloudflare D1CommercialCloudflare’s serverless SQL database that speaks SQLite, runs each database inside a Durable Object, and bills by rows read and written rather than by the hour.CloudNativePGOpen sourceA Kubernetes operator that runs highly available PostgreSQL with no Patroni and no StatefulSets: the Kubernetes API itself holds the cluster’s state.CockroachDBCommercialA distributed SQL database that speaks the PostgreSQL wire protocol, replicates every slice of data with Raft and keeps transactions serializable across nodes and regions.MySQLOpen sourceThe long-standing open-source SQL database behind much of the web, built on the InnoDB engine.NeonCommercialServerless Postgres that keeps data in a shared storage layer, so compute can scale to zero and a branch of the database is a pointer, not a copy.PatroniOpen sourceA Python agent beside each Postgres server that elects one primary through etcd, Consul, ZooKeeper or Kubernetes, and fails over when it dies.Percona OperatorsOpen sourceFour free Kubernetes operators from Percona that run its builds of MySQL, PostgreSQL and MongoDB, each using the engine’s own replication and backup tools.PGO (Crunchy Postgres for Kubernetes)Open sourceCrunchy Data’s Postgres operator: Patroni in every pod for failover, pgBackRest for backups, and a declarative `PostgresCluster` resource on top.PlanetScaleCommercialManaged MySQL on Vitess and managed Postgres, with schema changes shipped through branches and deploy requests.PostgreSQLOpen sourceThe open-source relational database that most new projects start on, extensible enough to grow into many niches.ScramDBCommercialA brand-new, closed-source database that speaks the PostgreSQL protocol and promises transactions, analytics and vector search on one copy of the data; first builds shipped in summer 2026.SQLiteOpen sourceA complete SQL database in a library and a single file, with no server to run.TursoOpen sourceA from-scratch Rust rewrite of SQLite with concurrent writes, plus libSQL (an open-contribution SQLite fork) and Turso Cloud, which hosts many small databases on S3.VitessOpen sourceA proxy layer, born at YouTube, that shards MySQL across many servers while the app still sees one database.

How it fits with other technologies

A primary and a read-only node sharing one data directory, in the open-source PolarDB for PostgreSQL 17 image (run on PolarDB 17.11.1.0) · bash

docker run -d --name polardb --cap-add=SYS_PTRACE --privileged=true \
  polardb/polardb_pg_local_instance:17
docker exec -it polardb bash

# the primary listens on 5432, a read-only node on 5433
psql -p 5432 -c "CREATE TABLE orders (id int PRIMARY KEY, total numeric);" \
             -c "INSERT INTO orders SELECT g, g * 1.5 FROM generate_series(1, 100000) g;"

psql -p 5433 -c "SELECT pg_is_in_recovery(), count(*), sum(total) FROM orders;"
#  pg_is_in_recovery | count  |     sum
# -------------------+--------+--------------
#  t                 | 100000 | 7500075000.0

psql -p 5433 -c "INSERT INTO orders VALUES (0, 0);"
# ERROR:  cannot execute INSERT in a read-only transaction

psql -p 5432 -Atc "SHOW polar_datadir;"
# file-dio:///var/polardb/shared_datadir

du -sh /var/polardb/*     # each node keeps only local state; the data is shared
# 468K  /var/polardb/primary_datadir
# 456K  /var/polardb/replica_datadir1
# 854M  /var/polardb/shared_datadir

The family#

PolarDB for MySQL, Cluster Edition

one writer

One primary and up to 15 read-only nodes on shared PolarStore storage, with failover in under 10 seconds.

PolarDB for MySQL, Multi-master Cluster (Limitless)

many writers

Every node accepts reads and writes; Alibaba says write throughput grows linearly with nodes, and switchover takes 5 to 10 seconds.

PolarDB for MySQL, Standard Edition

lower cost

The same kernel on ordinary ECS instances, with ESSD cloud disks or PolarStore underneath, on x86 or Arm (Yitian 710).

PolarDB for PostgreSQL

PostgreSQL 11 to 17

Shared-storage clusters like the MySQL edition, plus a distributed form with compute and data nodes for manual sharding below 1 PB.

PolarDB for PostgreSQL (Compatible with Oracle)

Oracle migration

Accepts Oracle stored procedures, data types, SQL and built-in functions with minimal changes.

PolarDB for Xscale (PolarDB-X)

distributed

Shared-nothing, MySQL-compatible, Paxos replication and TSO transactions; a Standard Edition upgrades in place to a distributed Enterprise Edition.

Serverless compute list prices#

PolarDB for MySQL serverless clusters, billed per PCU-second and settled hourly. Checked 11 October 2026 (page updated July 2026).

RegionEnterprise EditionStandard Edition
US (Virginia)$0.0794 per PCU-hour$0.07 per PCU-hour
Germany (Frankfurt)$0.1038 per PCU-hour$0.0915 per PCU-hour
Singapore$0.1160 per PCU-hour$0.1023 per PCU-hour
Chinese mainland$0.0611 per PCU-hour$0.0538 per PCU-hour
  • 1 PCU is about 1 core and 2 GB of memory; serverless clusters scale in steps of 0.5 PCU.
  • Storage is billed separately. PSL4 with a hot-standby storage cluster in a second zone: $0.000542 per GB-hour pay-as-you-go or $0.2569 per GB-month on subscription in Hong Kong and international regions, $0.000486 per GB-hour or $0.2308 per GB-month in the Chinese mainland. Single-zone storage costs half as much.

See it in a real stack

No architecture uses Alibaba Cloud PolarDB yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2018–2026

Before PolarDB, high-performance cloud databases usually ran on an instance store: a VM with a local SSD. Alibaba’s engineers listed the problems. Capacity was limited, the database had to replicate itself to survive a disk failure, the general-purpose file system’s kernel overhead wasted the speed of RDMA and NVMe, and there was no way for several database nodes to share one copy of the data. Open-source distributed file systems such as HDFS and Ceph had much higher latency than local disks.

Wei Cao and colleagues at Alibaba built PolarFS, a distributed file system with its network and I/O stacks in user space, and on top of it POLARDB, modified from AliSQL, Alibaba’s fork of MySQL and InnoDB. The primary and the read-only nodes share redo log and data files in the same directory on PolarFS. They described it at VLDB 2018, when it had recently become available on Alibaba Cloud, and Alibaba Cloud released PolarDB outside China in July 2019.

The family then widened: PostgreSQL and Oracle-compatible editions, and PolarDB-X, which started as a way past the scaling limits of the core transaction system behind Tmall’s Double Eleven sales. Alibaba open-sourced PolarDB for PostgreSQL under Apache 2.0 in 2021 and PolarDB-X that November. Research kept going: PolarDB Serverless (SIGMOD 2021) pulled memory into a remote pool and recovered from failures 5.3 times faster, PolarDB-IMCI (SIGMOD 2023) added a column index that sped up TPC-H queries by up to 149 times with under 5% impact on transactions, and multi-primary and CXL memory-pool designs won SIGMOD’s best industrial paper awards in 2024 and 2025.

Why it’s built this way

Shared storage instead of a full copy per replica
Instance stores capped capacity and couldn’t survive a disk failure without the database replicating itself. With one copy on a storage cluster, compute and storage scale separately, compute nodes hold no persistent state, and readers can be added without copying data.
A user-space I/O path on RDMA and NVMe
Research the team cited found about 20,000 instructions executed to complete a single 4 KB I/O through the legacy kernel stack. PolarFS is compiled into the database process and bypasses the kernel, which brought its write latency close to a local file system on SSD.
ParallelRaft instead of Raft
Raft’s strict ordering held back I/O on hardware with latencies of tens of microseconds. Databases tolerate I/O completing out of order, so ParallelRaft acknowledges, commits and applies log entries out of order while keeping replicas consistent.
Start from MySQL rather than a new engine
Modifying AliSQL, Alibaba’s existing MySQL fork, meant the new service kept MySQL’s protocol and SQL. Alibaba still sells PolarDB on moving MySQL 5.6, 5.7 and 8.0 databases without code changes.

Timeline

  1. 2018

    PolarFS and POLARDB described at VLDB

  2. 2019

    Released outside China in July

  3. 2021

    PolarDB Serverless at SIGMOD; PolarDB for PostgreSQL open-sourced (Apache 2.0), PolarDB-X in November

  4. 2023

    PolarDB-IMCI column index presented at SIGMOD

  5. 2024

    PolarDB-MP, a multi-primary design on disaggregated shared memory, wins SIGMOD’s best industrial paper

  6. 2025

    A CXL-based memory pool wins SIGMOD’s best industrial paper again

  7. 2026

    Open-source PolarDB for PostgreSQL 17.11.1.0 released in August

Sources16 · facts checked 11 October 2026

Improve this page on GitHub