One of the more novel differences between Aurora and other relational databases is how it pushes redo processing to a multi-tenant scale-out storage service, purpose-built for Aurora. Doing so reduces networking traffic, avoids checkpoints and crash recovery, enables failovers to replicas without loss of data, and enables fault-tolerant storage that heals without database involvement. Traditional implementations that leverage distributed storage would use distributed consensus algorithms for commits, reads, replication, and membership changes and amplify cost of underlying storage. In this paper, we describe how Aurora avoids distributed consensus under most circumstances by establishing invariants and leveraging local transient state. Doing so improves performance, reduces variability, and lowers costs.

Streaming replicas—and by extension, base backups—can become dangerously broken when the source and target machines run slightly different versions of glibc. Particularly, differences in strcoll and strcoll_l leave "corrupt" indexes on the slave. These indexes are sorted out of order with respect to the strcoll running on the slave. Because postgres is unaware of the discrepancy is uses these "corrupt" indexes to perform merge joins; merges rely heavily on the assumption that the indexes are sorted and this causes all the results of the join past the first poison pill entry to not be returned. Additionally, if the slave becomes master, the "corrupt" indexes will in cases be unable to enforce uniqueness, but quietly allow duplicate values.

Another core problem we’ve faced is one of the fundamental features of MongoDB (or any other schemaless storage engine): the lack of a schema. The lack of a schema may sound interesting, and in some cases it can certainly have its benefits. However, for many the usage of a schemaless storage engine leads to the problem of implicit schemas. These schemas aren’t defined by your storage engine but instead are defined based on application behaviour and expectations.

Since I am not ready to assert that there’s a distributed system I know to be better and simpler than eventually consistent datastores, and since I certainly know that InnoDB’s MVCC implementation is full of complexities, for right now I am probably in the same position most of my readers are: the two viable choices seem to be single-node MVCC and multi-node eventual consistency. And I don’t think MVCC is the simpler paradigm of the two.