Ordering and time

The older edit wins

A laptop clock is five minutes fast. It edits a title, then a phone edits the same title after receiving that change. A last-write-wins rule using device time can keep the laptop's older title. The timestamps look precise, but they do not prove the order people acted in. Clock sync reduces error; it does not remove uncertainty or pauses.

Order only what you know

Make edits with two drifting clocks. Receive the laptop history before a phone edit to create a causal dependency. Compare timestamp selection with vector-clock conflict detection.

Worked example: concurrent edits

Increment both devices, merge by component maximum, then merge again. Counter state can converge without treating repeated snapshots as new increments. Arbitrary text still needs a conflict policy.

Convergence is not every invariant

Convergence does not guarantee that the merged result preserves a business invariant. Concurrent adds, removes and text edits need deliberate rules and enough history to explain conflicts. A stock decrement may require coordination or preallocated rights. Prefer per-key order for dependent commands; independent counters or different users often need less coordination. If a requirement says newest, ask newest by whose observation.

Pick the invariant before the merge rule

• Independent counters: Merge per-participant maxima without adding a repeated state twice.
• Concurrent text edits: Keep operation identities and a deliberate conflict policy.
• One remaining seat: Coordinate or allocate rights before accepting competing decrements.

Key numbers

Clock error

≈ 300 seconds. A five-minute fast device can beat later wall-clock writes

Vector payload

≈ 800 bytes. 100 participants × 8-byte counters, before identities

Merged counter

7. max([4,0],[0,3]) = [4,3], then sum

Check your reasoning

Think through these questions before opening the lesson. In the game you can test the idea and compare your reasoning.

• Lamport(A) is smaller than Lamport(B). What can the counters alone establish?
• Why not use last-write-wins for notes? Follow-up: users rarely edit at the same time.
• Vectors [3,1] and [2,2] describe edits to one field. How should a causal merge classify them?
• Do we need one global event order? Follow-up: payments and click counters share the stream.
• Two disconnected sellers share a convergent sold-ticket counter but have no allocated rights. What prevents overselling?