Kyle Kingsbury's Jepsen project has tested almost every distributed database against partition tolerance and consistency claims since 2013. The lessons aren't 'database X is broken' — they're patterns that show up over and over. Knowing them helps you read marketing claims with the appropriate skepticism.
Default consistency isn't what marketing says
Many DBs default to lower isolation than implied. Mongo's 'majority' write concern isn't 'majority' read concern by default. Postgres's default is read committed, not serializable. Always check the actual isolation level you got, not what the home page implies.
Partition tolerance is partial in practice
Some DBs claim PA — they actually go inconsistent under specific partition patterns (asymmetric partitions, gray failures, slow networks). Jepsen finds these by constructing adversarial scenarios.
Clocks lie even in consensus systems
Several DBs that 'use logical clocks' actually fall back to wall clocks for tie-breaks. Skew between nodes (NTP issues, leap seconds) causes the rare-but-real incident. CockroachDB's TrueTime-like uncertainty intervals are a response.
Lost updates are the most common failure
Read-modify-write without proper isolation or conditional update loses concurrent updates. Even databases marketed as ACID can lose updates under default isolation. Test: increment a counter from N concurrent clients; expect N increments.
The lesson is procedural
Test your DB against your workload at partition. Use the actual consistency level you'll use in prod. Read Jepsen reports for the DB you chose. Treat the marketing page as marketing.