Does local serializability imply global serializability?

No. Local serialization orders may be incompatible.

Is two-phase commit the same property?

No. It is an atomic-commit protocol; global ordering additionally depends on concurrency-control behavior.

Must serializable transactions execute one at a time?

No. They can overlap if the resulting history is equivalent to an allowed serial execution.

AI Glossary / Data engineering and architecture / Intermediate

Global serializability

Definition

Global serializability is the property that transactions spanning several databases or resources have an overall effect equivalent to some serial execution of those transactions across the whole system. It extends serializability reasoning beyond one local database. Each participant being locally serializable is not, by itself, sufficient to guarantee a compatible global order.

What it means

A distributed transaction may touch data in more than one system. Each system can order conflicting operations in a locally valid way, yet those orders can disagree when considered together. Global serializability requires the combined history to fit one serial explanation under the relevant model. This is different from atomic commitment, which coordinates whether a transaction commits everywhere or aborts. Both properties can matter, but they solve different parts of transaction correctness.

How it works

A distributed concurrency-control design must prevent incompatible dependencies across participants. Possible approaches coordinate locks, timestamps, ordering information or other protocol-specific constraints. Commitment ordering is one researched approach whose guarantees depend on its stated assumptions and integration with atomic commitment. In conflict-serializability reasoning, a global precedence graph combines conflict edges from the participating resources; a cycle shows that no consistent serial order exists for that conflict model. Real multiversion and higher-level operation models require their own precise definitions.

Three practical examples

A transaction updates related records in two databases and must preserve a system-wide business invariant under concurrent activity.
A federated information system coordinates operations across autonomous database managers with different local concurrency mechanisms.
An AI artifact registry updates model metadata and associated release records across transactional resources, where the required correctness property must be defined rather than assumed from separate local commits.

Think of it this way

Think of two offices scheduling the same pair of jobs. One requires job A before job B, while the other requires B before A. Each local list looks orderly, but they cannot both describe one shared sequence. The analogy is limited because database dependencies arise from specific reads and writes, and distributed protocols must also handle aborts, failures and recovery.

Worked example

Consider transactions T1 and T2, both writing values at resources X and Y. At X, T1 writes 1 before T2 writes 2, leaving X equal to 2. At Y, T2 writes 2 before T1 writes 1, leaving Y equal to 1. If both transactions are treated as completed, the result (X=2, Y=1) is not the result of either serial order: T1 then T2 would leave both at 2, while T2 then T1 would leave both at 1. The example isolates incompatible local write orders; it is not a complete executable schedule for a particular database protocol.

Illustrative case study

Fictional case: Cedar Exchange integrates two transactional services. The team initially checks that each database uses serializable transactions, then analyses cross-service operation histories. It identifies a path where the services can impose opposing orders and revises the design to use a documented coordination mechanism. It also separates the atomic-commit test from the ordering test. This fictional case illustrates a reasoning gap, not a claim that every pair of serializable databases will necessarily fail.

Common misconceptions and mistakes

Do not equate global serializability with every participant committing successfully. Atomicity does not determine whether concurrent transactions have a serial explanation. Do not assume a message queue or two-phase commit alone supplies all required isolation guarantees. Serializability also differs from strict serializability, which adds a real-time ordering requirement. Finally, a distributed business workflow made of compensating steps may intentionally provide different guarantees from one global database transaction; describe those guarantees accurately.

Business use

Businesses need this concept when invariants cross transactional boundaries. First identify which operations must behave as one transaction and which weaker outcomes the application can tolerate. Evaluate the complete protocol, including retries, failures and coordination cost. For AI platforms, avoid using the word transaction loosely for a sequence of unrelated file and database writes. If a workflow does not provide global serializability, its consistency, reconciliation and recovery rules should still be explicit and testable.

Check your understanding

Resource X orders T1 before T2 while resource Y orders T2 before T1. Can both local orders always fit one global serial order?

Answer: No. These opposing dependencies form a cycle in the simple conflict model and require resolution.

Related terms

Distributed transaction
Commitment ordering
serializability
atomic commitment
two-phase commit
precedence graph
strict serializability

Learn next

Draw two transactions touching two resources and list the local conflict edges. Combine them into one graph and check for a cycle. Then identify which documented mechanism in a real design prevents or resolves the incompatible history.

Sources

Primary research references: Yoav Raz, Serializability by Commitment Ordering, https://www.sciencedirect.com/science/article/pii/0020019094900051 ; The Principle of Commitment Ordering, https://citeseerx.ist.psu.edu/document?doi=a33371ec053b61652f41d18e86dae07fe04a616b&repid=rep1&type=pdf . Original-paper abstracts and indexed excerpts consulted. Protocol theorems are not restated without their assumptions; full proof and editorial review remain pending.

Previous: Global optimization Next: Glottal