High upon the isolated crag of Cloud Ridge stood a lone beacon tower, weathered by endless gale and cloaked in drifting mountain mist. Below the ridge lay a treacherous pass where bandits and enemy scouts routinely crept through the foggy defiles under cover of night.

The outpost was manned by only two watchmen: seasoned Old Zhao and young Lu. Between them hung the ironclad edict of the garrison commander: “The winds over the pass are treacherous and the beacon must never stand unguarded. Between the two of you, at least one must remain awake on the parapet at all times.”

On the stone wall of the watchroom hung two carved wooden tablets: one inscribed with “Zhao” and the other with “Lu”. Beneath each tablet sat slots fitted with bronze markers engraved with “Awake” or “Resting”. To keep the men from quarreling or snatching markers at the same instant, the garrison had long ago established an orderly ritual: at each hourly chime of the bronze bell, a runner would take a fresh charcoal rubbing of the wall board and hand one copy to each watchman.

Whoever wished to step down from the windy platform to rest had to consult the rubbing in his palm first: if the rubbing showed the other watchman marked as “Awake”, the sacred invariant held true, and he was permitted to swap his bronze marker to “Resting” and slide it into the iron confirmation chute. So long as no two men wrestled for the very same marker slot, the watch transitioned smoothly. For years, this practice had kept the frontier secure without a single hitch.

On a bitter night at the close of autumn, rain lashed the mountain peak like knives. Both men had endured ten grueling hours in the biting downpour, their eyelids heavy as wet stones.

At midnight, the bronze bell sounded its deep chime. The runner delivered the freshly inked charcoal rubbings into the trembling hands of Old Zhao and young Lu.

Huddled against the stone parapet at the top of the tower, Old Zhao pulled out the rubbing and inspected it by the dim light of his oil lantern: Zhao was “Awake”; Lu was “Awake”.

Old Zhao reasoned carefully: the law demanded that at least one guard remain awake. Lu was awake on duty. If Zhao stepped down to sleep beside the brazier for two hours, the rule would still hold: Lu was watching the gorge. “The rule is satisfied. It is entirely safe.” With a sigh of relief, Old Zhao swapped his marker to “Resting”, slid it into his designated slot in the chute, and retreated to the warm quarters to collapse into deep slumber.

At that very same moment, on the sheltered lower walkway, young Lu examined his own rubbing from the exact same midnight bell. The black ink was identical: Zhao was “Awake”; Lu was “Awake”.

Young Lu’s boots were soaked through, and the chill bit deep into his marrow. Gazing at the slip, he too breathed a sigh of relief: “Old Zhao is veteran and wise. Since he is awake and watching the beacon on the top platform, I can return to the hearth room, dry my clothes, and rest my limbs. As long as Zhao is up there, the law is unbroken.” Lu dutifully changed his own marker to “Resting” and dropped it into his slot.

The iron mechanisms inside the chute clicked smoothly into place. Zhao had modified only Zhao’s marker; Lu had modified only Lu’s marker. Neither man touched the other’s slot; neither attempted to overwrite the same bronze piece. Both slips locked home without collision.

Yet on the parapet above, the lashing rain snuffed out the last dying embers of the watch fires.

In pitch blackness and deafening rain, the tower stood entirely abandoned. In the fourth watch, a troop of masked horsemen slipped through the unguarded pass without raising a whisper of alarm.

At dawn, the inspecting magistrate arrived with cavalry only to find the beacon drum stone-cold and both watchmen wrapped in heavy quilts, snoring peacefully. Dragged out into the morning fog, the two men cried out in disbelief, thrusting forward their charcoal rubbings and stamped receipts: “Excellency, look! When I checked the record before submitting my marker, the other man was wide awake! Every rule was observed to the letter! How could the tower be empty?!”

The magistrate cast the rubbings onto the stones: “Each of you saw a true picture of the world, and neither of you broke a rule on your own marker. Yet because each of you made decisions on your own snapshot and altered different markers, Cloud Ridge became an empty gate.”

By now, you have probably recognized it: this is the write skew anomaly under Snapshot Isolation in database concurrency control.

What it is

In database systems, Snapshot Isolation (SI) is a widely used, high-performance concurrency control model. Under Snapshot Isolation, every transaction reads data from a consistent snapshot taken when the transaction begins—just like the freshly inked charcoal rubbing. As long as concurrent transactions do not attempt to write to the exact same row or data item, no write-write conflict is triggered (first-committer-wins), and both transactions are permitted to commit.

However, whenever an application’s integrity relies on a global invariant spanning multiple distinct rows, Snapshot Isolation breaks down.

The classic write skew pattern unfolds precisely like Cloud Ridge:

1. The invariant is: count(awake_guards) >= 1.
2. Transaction A (Old Zhao) reads the snapshot, observes that two guards are awake, and concludes it is valid to update his own status to resting (modifying row A), as the invariant will remain satisfied.
3. Transaction B (young Lu) concurrently reads the same snapshot, also sees that two guards are awake, and concludes it is valid to update his own status to resting (modifying row B).
4. Both transactions commit. Because transaction A touched only row A and transaction B touched only row B, their write sets are disjoint. Under standard Snapshot Isolation, no row-level write conflict exists, and both transactions commit cleanly.
5. Final state: both guards are resting, count(awake_guards) = 0, and the multi-row invariant is completely shattered.

Because the two concurrent writes “skew” away from each other onto different records, they slip right past row-level conflict detection—hence the name Write Skew.

Why it matters

Many developers assume that “Snapshot Isolation” or SQL’s “Repeatable Read” (as implemented in PostgreSQL, MySQL InnoDB, CockroachDB, or Spanner) protects against all race conditions. Write skew is the primary blind spot where that assumption fails:

• On-call rotations: A hospital schedule requires at least one physician on call per department. Two doctors concurrently request leave on their phones; both see the other active, both requests are approved automatically, leaving the department unstaffed.
• Overdraft protection across accounts: A user has linked checking and savings accounts with the rule that checking_balance + savings_balance >= 0. Concurrent debits against checking and savings, each checking the aggregate balance from an initial snapshot, will overdraw the total balance.
• Meeting room & resource reservations: Checking availability in a timeslot and then inserting a new booking row without a lock allows two concurrent transactions to book overlapping reservations for the same physical room.

There are three primary remedies for write skew:

1. Serializable Snapshot Isolation (SSI): Implemented in modern engines such as PostgreSQL’s SERIALIZABLE level. SSI tracks read-write dependency cycles (rw-antidependencies) during transaction execution and automatically aborts one of the conflicting transactions before a violation can occur.
2. Explicit pessimistic locks (SELECT ... FOR UPDATE): Forcing a lock on all rows contributing to the invariant forces concurrent transactions to serialize, ensuring the second transaction sees the up-to-date state.
3. Materialized conflicts: Artificially creating a shared “guard lock” record that any transaction modifying the schedule must touch and update, converting a multi-row semantic conflict into a standard, detectable row-level write-write conflict.

Metaphor mapping

• Watchtower invariant (at least one guard awake) → Global integrity invariant spanning multiple rows
• Charcoal rubbings taken at the hourly bell → Transaction snapshot view taken at start time
• Zhao and Lu inspecting their rubbings → Concurrent transactions reading overlapping data
• Zhao modifying Zhao’s marker, Lu modifying Lu’s marker → Concurrent transactions writing to disjoint rows
• Bronze markers sliding into the chute without collision → Row-level conflict check passing (no write-write conflict)
• Unguarded parapet and captured pass → The invariant-violating Write Skew Anomaly
• Contending for a single shared bell mechanism → Explicit locks (SELECT FOR UPDATE) or materialized conflict records