Case study

Design a Unique ID Generator

Generate globally unique, roughly sortable 64-bit IDs at 10K+/sec without coordination per ID.

SnowflakeUUIDdistributed

Requirements

Functional: generate unique ID; optionally expose batch generate
Properties: unique across all datacenters; roughly time-ordered (sortable)
Non-functional: 10K IDs/sec per machine; latency < 1ms local
No central DB round-trip per ID (bottleneck)
IDs numeric or fixed-length string (base62)

Compare to UUID v4

UUID v4 is random — not sortable and poor for B-tree indexes. Snowflake/truncated timestamp IDs give locality that improves DB insert performance.

Back-of-envelope Estimation

Metric	Value	Notes
Global ID rate	1M/sec peak	Twitter-scale order of magnitude
Per machine	10K/sec	100 machines suffice with headroom
ID size	64 bits (8 bytes)	Fits in BIGINT; 18 decimal digits max
Clock drift budget	< 5 ms	NTP sync required per host

API Design

Endpoint	Description
GET /v1/id	Returns { id: 1234567890123456789 }
GET /v1/id/batch?count=100	Returns array; max 1000 per request
Internal: allocate_worker_id	ZooKeeper/etcd lease assigns machine ID 0–1023

Snowflake Bit Layout

Field	Bits	Range
Unused/sign	1	0
Timestamp (ms since epoch)	41	~69 years
Datacenter + machine ID	10	1024 machines
Sequence	12	4096 IDs/ms per machine

Per millisecond: increment sequence. If sequence overflows, spin until next ms. On clock backward jump, wait or error — never reuse timestamp+sequence pairs.

High-level Design

ID service deployment

App servers ──► ID Generator lib (embedded) ──► local clock + worker_id
                    │
                    └── optional central service for worker_id lease only

ZooKeeper/etcd: /workers/{dc}/{host} → worker_id (ephemeral node)

Prefer embedded library (Twitter Snowflake style) over RPC per ID — eliminates network hop. Central service only coordinates worker_id assignment on startup.

Deep Dive: Alternatives

Approach	Pros	Cons
DB auto-increment	Simple, strict order	Single-writer bottleneck; hard to shard
UUID v4	No coordination	Random; 128 bits; index fragmentation
Snowflake	Sortable; high throughput	Clock dependency; machine ID management
Redis INCR	Easy	Single point of failure; network per ID

Bottlenecks & Edge Cases

Clock skew backward — refuse to generate until caught up; metric clock_rollback_events
Worker ID exhaustion (1024 machines) — extend bits or hierarchical IDs per service
Leap seconds / VM pause — sequence buffer absorbs brief stalls; alert on NTP drift > 1s
Hot partition in DB if IDs used as sole shard key — combine with secondary sharding key

Failure Modes & Monitoring

Metric	Alert threshold
ids_generated/sec per host	Approaching 4096/ms sustained
clock_rollback_count	> 0 per minute
worker_id lease renewals failed	Any failure
duplicate_id_detected	Critical — should never happen

Testing uniqueness

Chaos-test clock jumps and worker restarts. Run collision detector in staging comparing ID sets across nodes.

System Design