Redis as In-Memory Database

While commonly used as a cache, Redis can serve as a primary database for workloads where speed matters more than ACID guarantees and where the dataset fits in RAM. Understanding its persistence and durability characteristics is essential for senior engineers.

---

Redis as Primary DB vs Cache

Dimension	Redis as Cache	Redis as Primary DB
Data loss tolerance	Acceptable	Not acceptable
Dataset size	Can exceed RAM (use eviction)	Must fit in RAM
Persistence	Optional	Required (RDB + AOF)
Consistency	Eventual	Depends on config
Recovery time	Fast (warm cache)	Depends on AOF size
Use case	Speed layer on top of real DB	Sessions, leaderboards, real-time data

---

Persistence Mechanisms

1. RDB (Redis Database Snapshot)

Creates a point-in-time binary snapshot of the entire dataset via fork().

# redis.conf
save 3600 1      # Snapshot if 1+ change in last 3600s
save 300 100     # Snapshot if 100+ changes in last 300s
save 60 10000    # Snapshot if 10000+ changes in last 60s

# Manual trigger
BGSAVE           # Forks Redis process, writes snapshot in background
LASTSAVE         # Unix timestamp of last successful save

# Disable all saving
save ""          # Pure cache mode — no persistence

How RDB fork() works:

Main Redis Process (parent)
    ↓ fork()
Child Process (snapshot writer)
    ↓
Writes binary .rdb file to disk

Main process continues serving requests.
Copy-on-Write (CoW): modified pages are copied when written.
Both processes see the same memory initially.

RDB trade-offs:

Pro	Con
Compact single file	Data loss between snapshots
Fast DB load on restart	fork() can cause latency spike (large heaps)
Good for backups	Not suitable for strict durability
Minimal IO during operation	10GB heap → 100–200ms latency spike on fork

Monitoring RDB fork latency:

INFO latencystats
LATENCY HISTORY fork
# Or check: rdb_last_bgsave_time_sec

2. AOF (Append-Only File)

Logs every write command to disk. On restart, Redis replays the AOF to rebuild the dataset.

# redis.conf
appendonly yes
appendfsync everysec    # Recommended — sync to disk every second
# appendfsync always    # Sync after every write (slowest, most durable)
# appendfsync no        # OS decides when to flush (fastest, least safe)

AOF fsync modes:

Mode	Durability	Performance	Risk
always	Max (~0 data loss)	Slow (disk write per command)	Low throughput
everysec	~1 second of data loss	Good	Recommended
no	OS-controlled	Fastest	Up to kernel buffer loss on crash

AOF Rewrite (compaction):

# AOF grows indefinitely — rewrite compresses it
BGREWRITEAOF   # Background AOF rewrite

# Auto-rewrite when AOF grows
auto-aof-rewrite-percentage 100   # Rewrite when AOF doubles in size
auto-aof-rewrite-min-size 64mb   # Minimum size before rewrite

During rewrite, Redis forks and writes a new AOF from current in-memory state. Parent continues appending to the old AOF; differences are merged when the child completes.

AOF trade-offs:

Pro	Con
Near-real-time durability	Larger file than RDB
Can reconstruct to exact state	Slower restart (replay all commands)
Human-readable format	AOF rewrite causes another fork() spike
everysec = at most 1s data loss	Old commands not compacted until rewrite

3. RDB + AOF (Recommended for Production)

Use both for the best of both worlds:

appendonly yes
appendfsync everysec
save 3600 1
save 300 100

# On restart: Redis prefers AOF (more complete); RDB used as backup
aof-use-rdb-preamble yes   # Redis 4.0+: AOF starts with RDB snapshot header → faster rewrite

Recovery scenarios:

Failure	RDB only	AOF only	Both
Server crash	Lose since last snapshot	Lose up to 1s	Lose up to 1s
Corrupted AOF	N/A	Need redis-check-aof repair	RDB as fallback
Disk failure	Lose everything	Lose everything	Same

---

Data Modeling for Redis as Primary DB

Use Case: User Sessions

# One hash per session (rich structure, partial field update)
HSET session:abc123 userId "1234" role "admin" expiresAt "1700000000"
EXPIRE session:abc123 3600

# Index by userId (for "show all sessions for user")
SADD user:1234:sessions "abc123"

Use Case: Real-Time Leaderboard

# Sorted Set — atomic, O(log N) updates
ZADD game:leaderboard 15000 "alice"
ZADD game:leaderboard 12500 "bob"
ZINCRBY game:leaderboard 500 "alice"          # Atomic score increment
ZREVRANK game:leaderboard "alice"              # Rank (0-indexed) → 0 (first place)
ZREVRANGEBYSCORE game:leaderboard +inf -inf WITHSCORES LIMIT 0 10   # Top 10

Use Case: Rate Limiting (Primary Store)

-- Lua script: atomic sliding window rate limit
local key = KEYS[1]
local limit = tonumber(ARGV[1])
local window = tonumber(ARGV[2])
local now = tonumber(ARGV[3])

redis.call('ZREMRANGEBYSCORE', key, 0, now - window * 1000)
local count = redis.call('ZCARD', key)
if count < limit then
    redis.call('ZADD', key, now, now .. '-' .. math.random())
    redis.call('PEXPIRE', key, window * 1000)
    return 1  -- allowed
end
return 0  -- blocked

Use Case: Distributed Lock (Redlock)

# Single-node lock (sufficient for most cases)
SET lock:resource "owner-uuid" NX EX 30
# NX: only set if not exists
# EX: auto-release after 30s (prevents deadlock if owner crashes)

# Release lock (Lua for atomicity — don't release if owner changed)
if redis.call('GET', KEYS[1]) == ARGV[1] then
    return redis.call('DEL', KEYS[1])
end
return 0

Redlock (multi-node): For stronger guarantees, acquire lock on N/2+1 nodes:

# Redlock algorithm (5 Redis nodes)
nodes = [redis1, redis2, redis3, redis4, redis5]
acquired = sum(1 for r in nodes
               if r.set("lock:key", token, nx=True, ex=30))
if acquired >= 3:  # Majority
    return True  # Lock acquired

Clock drift caveat: Redlock assumes clocks don't drift significantly. Martin Kleppman's critique: use fencing tokens (monotonic counter) for true linearizability.

---

Durability Guarantees by Configuration

Config	Max Data Loss	Throughput	Use Case
No persistence	100%	Highest	Pure cache
RDB only (60s)	Up to 60s	High	Cache with backup
AOF no	OS buffer	High	Balanced
AOF everysec	Up to 1s	Medium-High	Recommended primary
AOF always	~0	Low	Critical data
RDB + AOF everysec	Up to 1s	Medium-High	Production DB

---

Redis vs Traditional Databases

	Redis	PostgreSQL / MySQL	MongoDB
Storage	In-memory (optional disk)	Disk-first	Disk-first
Throughput	100K–1M+ ops/sec	1K–100K ops/sec	10K–100K ops/sec
Query language	Command-based	SQL	Query DSL
Joins	❌ (model differently)	✅	❌ (embed/link)
Transactions	Limited (MULTI/EXEC)	Full ACID	Multi-doc (limited)
Dataset size	Limited by RAM	Unlimited	Unlimited
Full-text search	Limited	pgvector, FTS	Atlas Search
Best for	Speed-critical, structured data	Complex queries, ACID	Flexible documents

---

When to Use Redis as Primary DB

Good fit:

• Sessions, tokens, temporary data with natural TTL
• Leaderboards, rankings, counters
• Rate limiting, quota tracking
• Short-lived job/task queues
• Real-time collaborative features (cursors, presence)
• Configuration/feature flags

Bad fit:

• Complex relational queries (joins, aggregations)
• Large datasets that don't fit in RAM
• Long-term audit logs (use Elasticsearch or cold storage)
• Financial transactions requiring ACID guarantees
• Full-text search (use Elasticsearch)