Do we fear the serializable isolation level more than we fear subtle bugs?
After answering these questions, we concluded that using weaker isolation levels by default is a form of premature optimization and evil. Consider using serializable isolation as the default unless your DBMS is CockroachDB or YDB, where serializable isolation is already the default.
Suppose a table with a single column named “column” contains strings “black” or “white”. One user wants to change all “white” colors to “black”, while another user concurrently tries to change “black” to “white”. In other words, there are two concurrent transactions:
What would be the result of these two transactions? Intuitively, all colors should either become black or white. But in database practice, the proper answer is “it depends on isolation level.”
While “Isolation” is supposed to originally mean “serializability”, there are weaker [isolation levels](https://en.wikipedia.org/wiki/Isolation_(database_systems) introduced as a trade-off between performance and safety:
