A Process phenomenon involves two operations executed by the same logical process. If the process executes one operation before another, we say they have a process dependency. If we say that each database session is a single process, process phenomena are session phenomena.

Process phenomena always involve some other relationship between the operations. For instance, a process might fail to observe a write it performed in an earlier transaction. The later transaction rw-depends on the earlier). Jepsen calls this G2-item-process, by analogy with G2-item.

For example, imagine a person with ADHD writes themselves detailed shopping lists. They write down that they intend to purchase milk at the store. When they arrive at the store they check their notes app, and (perhaps due to a bug) observe an earlier state of their shopping list. They return home without the milk they needed.

More formally, we say that given some phenomenon P defined in terms of a cycle in a dependency graph G, the phenomenon P-process is the corresponding cycle in the graph (G ∪ P), where P is the graph of process dependencies between transactions in G.

In single-object systems, Sequential Consistency prevent Process phenomena. The “Strong Session” variants of transactional consistency models generally differ from their non-strong counterparts by ruling out Process phenomena as well. For instance, Serializable proscribes G1c and G2: cycles involving write-write, write-read, and read-write dependencies. Strong Session Serializable goes further by proscribing G1c-process and G2-process: cycles involving write-write, write-read, read-write, and process dependencies. Similarly, Strong Session Snapshot Isolation strengthens Snapshot Isolation by ruling out G-nonadjacent-process.

Literature

Daudjee and Salem’s Lazy Database Replication with Ordering Guarantees introduced the concept of Strong Session Serializability. They partitioned transactions into sessions, and enforced ordering constraints only between transactions in a single session. Two years later, their Lazy Database Replication with Snapshot Isolation introduced Strong Session Snapshot Isolation, along similar lines. Daudjee and Salem’s formal definitions of sessions were arbitrary, but the intuition is clear: sessions are meant to capture operations performed by some single-threaded logical process.

For symmetry with Real-Time phenomena, Jepsen adopts Adya’s technique of defining additional dependency edges between transactions. We construct a graph identical to the Direct Serialization Graph, but with extra process edges that capture when a single process performs two transactions sequentially.

See Also

Real-Time phenomena are like Process phenomena, but across all processes, rather than just one.