March 16, 2026
Multi-master or multi-disaster?
Jepsen: MariaDB Galera Cluster 12.1.2
Jepsen says MariaDB’s cluster can lose data; vets say “workhorse,” critics cry “yikes”
TLDR: Jepsen’s new tests say MariaDB’s Galera Cluster can show data consistency issues despite docs promising no losses. Commenters split: some rely on Galera as a dependable workhorse for small workloads, while others call the problems fundamental and warn that retrofitting distributed guarantees is risky—and that matters when your data is money
Jepsen just threw shade at MariaDB’s Galera Cluster, calling out glossy promises like “no lost transactions” and “committed on all nodes,” and showing cases where money can seemingly appear or vanish in tests. The community went full popcorn mode. One reader demanded the headline say what everyone’s thinking: “Data loss cases…” Another pointed to the fine print: even in healthy setups, Jepsen reports stale reads and lost updates—translation: you might read yesterday’s truth after writing today’s changes. And those marketing lines about every server agreeing instantly? Jepsen says that’s not how it works. Think quorum votes instead of everyone holding hands.
But it’s not all pitchforks. A long-timer defended Galera as a “real workhorse” for low-traffic, multi-master needs—perfect for small orgs that want every server to accept changes. Others shrugged that this is what happens when you bolt a distributed system onto a tool that wasn’t born that way. The thread spawned memes about “multi-master, multi-disaster,” and one zinger dubbed it “Schrödinger’s transaction”—both committed and not, depending on the node. Drama aside, the stakes are real: if the docs overpromise and the lab tests say otherwise, teams might be banking on guarantees they don’t actually have. And that’s the kind of plot twist nobody wants in production
Key Points
- •Jepsen revisits MariaDB Galera Cluster, recalling 2015 findings that Galera lacked first-committer-wins despite Snapshot Isolation claims, enabling transactional anomalies.
- •Galera’s architecture: optimistic execution on any node; synchronous replication at commit; certification on other nodes based on written primary keys; conflict detection via seqno.
- •MariaDB documentation claims unanimous replication and no lost transactions, but Galera actually progresses with quorum and tolerates minority node failures.
- •Earlier Galera versions required manual recovery after quorum loss; newer versions can recover automatically.
- •Documentation provides limited clarity on true isolation levels; a note states Galera’s effective isolation lies between Serializable and Repeatable Read, ignoring tx_isolation settings.