Seeing is believing: a client-centric specification of database isolation, Crooks et al., PODC’17.
Last week we looked at Elle, which detects isolation anomalies by setting things up so that the inner workings of the database, in the form of the direct serialization graph (DSG), can be externally recovered. Today’s paper choice, ‘Seeing is believing’ also deals with the externally observable effects of a database, in this case the return values of read operations, but instead of doing this in order to detect isolation anomalies, Crooks et al. use this perspective to create new definitions of isolation levels.
It’s one of those ideas, that once it’s pointed out to you seems incredibly obvious (a hallmark of a great idea!). Isolation guarantees are a promise that a database makes to its clients. We should therefore define them in terms of effects visible to clients – as part of the specification of the external interface offered by the database. How the database internally fulfils that contract is of no concern to the client, so long as it does. And yet, until Crooks all the definitions, including Adya’s, have been based on implementation concerns!
In theory defining isolation levels Continue reading
