What does it mean for a concurrent schedule to be serialisable, and how does two-phase locking guarantee it?

The correctness criterion

Running transactions one at a time is obviously correct and unacceptably slow. Everything in concurrency control is an attempt to interleave operations while preserving the correctness that the serial execution would have given you, so the first thing to be precise about is what "correct" means here.

A schedule — a particular interleaving of the reads and writes of several transactions — is serialisable if its effect is equivalent to some serial execution of those same transactions. Not a specific one. If T1 and T2 run concurrently and the outcome matches either "T1 then T2" or "T2 then T1", the schedule is correct. The database is under no obligation to produce the order you expected, only to produce a result consistent with one of them.

That "some" is what candidates most often drop, and it is the whole point. It is why two concurrent transfers between accounts can be interleaved freely as long as the final balances match one of the two orderings, and why a system that insisted on a particular order would be needlessly serial.

Why conflicts are the only thing that matter

Two operations conflict when they belong to different transactions, touch the same item, and at least one is a write. Read-read does not conflict; the order is unobservable. Everything else does.

This gives a workable test. Draw a node per transaction and an edge from T1 to T2 whenever an operation of T1 conflicts with a later operation of T2 on the same item. That edge says T1 must precede T2 in any equivalent serial order. If the resulting graph is acyclic, a valid serial order exists — a topological sort of the graph produces one — and the schedule is conflict serialisable. A cycle means the schedule demands that T1 precede T2 and T2 precede T1, which no serial order can satisfy.

Conflict serialisability is slightly stricter than serialisability in general, since a schedule can accidentally produce a serialisable result through writes that cancel out. Databases use the conflict form because it is checkable from the operations alone, without knowing what the transactions mean.

What two-phase locking does

2PL is a protocol with one rule: a transaction may not acquire any lock after it has released one. Its lifetime therefore has a growing phase in which it only takes locks, and a shrinking phase in which it only releases them, with a single point between them where it holds the maximum set.

That one rule is sufficient to guarantee conflict serialisability, and the argument for why is short. Consider that maximum-holdings point — the lock point — for each transaction. Any conflict between two transactions is mediated by a lock on the item they both touch, so one must have released it before the other acquired it, which orders their lock points. Ordering the transactions by lock point therefore gives a serial order consistent with every conflict, and a consistent order cannot contain a cycle.

Note what this does not say. 2PL never inspects a dependency graph or checks for cycles. It is a local rule each transaction follows independently, and serialisability falls out as a property of the whole system. That is why it is implementable — a cycle check across all in-flight transactions on every operation would not be.

Strict 2PL, and why the extra restriction

Plain 2PL is serialisable but permits a nasty failure. A transaction can release a lock in its shrinking phase, another transaction reads the value it wrote, and then the first transaction aborts. The second has now read data that never officially existed — a dirty read — and to fix it you must abort the second transaction too, and anything that read from it, in a cascade.

Strict 2PL holds all exclusive locks until commit or abort, collapsing the shrinking phase into the moment the transaction ends. Nobody can observe an uncommitted write, so rollback affects only the transaction doing it. Essentially every locking database implements the strict variant, and when a system says "two-phase locking" this is almost always what is meant.

The price is deadlock, and it is not a bug

Because 2PL forbids acquiring after releasing, a transaction that needs a lock held by another must wait rather than release what it holds and retry. Two transactions that need each other's locks wait forever.

This is inherent to the protocol, not a flaw in a particular implementation. You can prevent it by requiring every transaction to acquire all locks up front, or to acquire them in a global order, but both require knowing the full access set in advance — which a query with a WHERE clause does not. So real systems detect instead: they maintain a wait-for graph, look for cycles, and abort a victim, or they simply time out. Either way the application must be prepared to retry, and a transaction that cannot be safely retried is an application-level design problem the database cannot solve for you.

The alternative lineage avoids the conflict rather than managing it. Multiversion concurrency control keeps old versions of rows, so a reader sees a consistent snapshot without taking any lock and never blocks a writer. That removes read-write deadlocks entirely, at the cost of storing versions and of an isolation level that is not quite serialisable — snapshot isolation still permits write skew, where two transactions each read a state, each write something that would have been invalid had it seen the other's write, and neither conflicts directly. Getting from there to genuine serialisability needs extra machinery, which is what PostgreSQL's serialisable isolation level adds on top of its snapshot machinery.

Serialisable means equivalent to some serial order, not to the one you had in mind. Two-phase locking earns that guarantee with a purely local rule, and deadlock is the bill.