When “now” is different everywhere

In distributed systems, time is a guess.
And guesses break correctness.

1️⃣ The False Assumption (Root of All Time Bugs)

“Time moves forward at the same rate everywhere.”

This is never true in distributed systems.

Reasons:

Clock drift (hardware)
NTP adjustments
VM pauses
GC pauses
Leap seconds
Network delays

Two machines disagreeing by milliseconds is normal.
By seconds is common.
By minutes happens in production.

2️⃣ What Is Clock Skew?

Clock Skew = difference between clocks on different machines.

Machine A: 10:00:01
Machine B: 09:59:58

Both are “correct” locally.

3️⃣ Why Time-Based Logic Is Dangerous

Time is often used to:

Order events
Enforce uniqueness
Expire data
Resolve conflicts
Detect staleness

Every one of these can break under skew.

4️⃣ Bug #1 — Time-Based Ordering Is Wrong

❌ Naive Event Ordering

events.sort((a, b) => a.timestamp - b.timestamp);

What you assume

Earlier timestamp → happened first

Reality

Event B may have happened after A but recorded earlier

Sorting by timestamp lies about causality.

5️⃣ Bug #2 — “Latest Write Wins” Loses Writes

❌ LWW Conflict Resolution

if (incoming.timestamp > stored.timestamp) {
  overwrite();
}

Failure Scenario

Node with fast clock overwrites correct data
Node with slow clock loses updates forever

Correctness depends on clock accuracy, which you don’t control.

6️⃣ Bug #3 — Time-Based IDs Are Not Unique

❌ Timestamp-based IDs

const id = Date.now();

Under concurrency:

Same millisecond
Same ID
Collision

Worse across machines:

Clock jumps backward → duplicates

7️⃣ Bug #4 — TTL & Expiry Bugs

❌ Local Time Expiry

if (Date.now() > expiresAt) {
  invalidate();
}

Failure Modes

Clock jumps forward → premature expiry
Clock jumps backward → never expires
Different nodes disagree on validity

8️⃣ Bug #5 — Timeouts That Aren’t Comparable

❌ Distributed Timeout Logic

if (Date.now() - start > timeout) {
  fail();
}

If start was recorded on another machine:

Negative elapsed time
Infinite waits
Premature failures

9️⃣ Why NTP Doesn’t “Fix” This

NTP:

Adjusts clocks gradually
Sometimes jumps time
Can go backward

NTP reduces skew.
It does not eliminate it.

1️⃣0️⃣ The Golden Rule of Time

Never use wall-clock time to establish ordering or correctness.

Use time only for:

Human display
Logging
Approximate expiry (with tolerance)

1️⃣1️⃣ Correct Tool #1 — Monotonic Clocks

Monotonic clocks:

Always move forward
Not affected by NTP
Local only

✅ Correct Timeout Measurement

const start = performance.now();
// work
if (performance.now() - start > timeoutMs) {
  fail();
}

Monotonic clocks are safe only locally.

1️⃣2️⃣ Correct Tool #2 — Versioning (Not Time)

Replace timestamps with:

Versions
Counters
CAS tokens

UPDATE doc
SET value = ?, version = version + 1
WHERE id = ? AND version = ?

Correctness without time.

1️⃣3️⃣ Correct Tool #3 — Logical Clocks

Lamport Clock (Simple)

counter++

Attach counter to events.

Guarantee:

Causal ordering (not real time)

1️⃣4️⃣ Correct Tool #4 — Vector Clocks (Advanced)

Track causality across nodes.

Use when:

Conflict resolution matters
Order matters
Strong correctness required

Cost:

Metadata size
Complexity

1️⃣5️⃣ Correct Tool #5 — Server-Assigned Time

If time is needed:

Assign it in one place
Accept bottleneck

INSERT INTO events (created_at)
VALUES (CURRENT_TIMESTAMP);

Centralized time is slow — but correct.

1️⃣6️⃣ Why “Time-Based Sharding” Is Risky

logs_2026_10
logs_2026_11

Clock skew →

Writes go to wrong shard
Reads miss data
Data loss illusions

Always add overlap or buffers.

1️⃣7️⃣ Production Rulebook (Hard-Won)

Never compare timestamps from different machines
Never rely on client time
Never order events by wall-clock time
Use monotonic clocks for durations
Use versions for correctness
Assume clocks go backward

📘 Clock Skew & Time-Based Bugs