Redis High Availability in Production: Sentinel, Cluster, and Kubernetes Operators

Redis HA Helm Values for k3s

# values-redis-ha.yaml — Helm values for Redis HA on k3s/Longhorn
redis:
  replicas: 3
  resources:
    requests:
      cpu: "2"
      memory: 8Gi
    limits:
      cpu: "4"
      memory: 16Gi
  storage:
    persistentVolumeClaim:
      metadata:
        name: redis-data
      spec:
        accessModes:
          - ReadWriteOnce
        resources:
          requests:
            storage: 50Gi
        storageClassName: longhorn
  customConfig:
    maxmemory: "6gb"
    maxmemory-policy: "allkeys-lru"
    save: "900 1 300 10 60 10000"
    appendonly: "yes"
    appendfsync: "everysec"
    aof-use-rdb-preamble: "yes"
    repl-backlog-size: "256mb"
    tcp-keepalive: "60"
    timeout: "300"
    hz: "10"
    activedefrag: "yes"
  exporter:
    enabled: true
    image: oliver006/redis_exporter:latest

sentinel:
  replicas: 3
  resources:
    requests:
      cpu: 100m
      memory: 128Mi
    limits:
      cpu: 200m
      memory: 256Mi
  customConfig:
    down-after-milliseconds: "5000"
    failover-timeout: "60000"
    parallel-syncs: "1"

affinity:
  podAntiAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      - labelSelector:
          matchExpressions:
            - key: app.kubernetes.io/component
              operator: In
              values:
                - redis
        topologyKey: kubernetes.io/hostname

Memory Management and Eviction Policies

Redis stores all data in memory, making memory management the most critical operational concern. When Redis reaches the configured maxmemory limit, it must decide what to do with incoming write commands. The eviction policy controls this behaviour.

# redis.conf — Memory management
maxmemory 6gb
maxmemory-policy allkeys-lru

# Available eviction policies:
# noeviction        — Return errors on writes when memory limit reached
# allkeys-lru       — Evict least recently used keys (general-purpose cache)
# allkeys-lfu       — Evict least frequently used keys (better for skewed access patterns)
# volatile-lru      — Evict LRU keys with TTL set
# volatile-lfu      — Evict LFU keys with TTL set
# volatile-ttl      — Evict keys with shortest TTL first
# allkeys-random    — Evict random keys
# volatile-random   — Evict random keys with TTL set

# Active defragmentation (Redis 4.0+)
activedefrag yes
active-defrag-enabled yes
active-defrag-ignore-bytes 100mb
active-defrag-threshold-lower 10
active-defrag-threshold-upper 100
active-defrag-cycle-min 1
active-defrag-cycle-max 25
active-defrag-max-scan-fields 1000

# Memory usage monitoring
# redis-cli INFO memory
# Key metrics:
#   used_memory           — Total bytes allocated by Redis
#   used_memory_rss       — Resident set size (OS-level memory)
#   mem_fragmentation_ratio — RSS / used_memory (should be close to 1.0)
#   maxmemory             — Configured memory limit
#   evicted_keys          — Total keys evicted due to maxmemory

For HA deployments, set maxmemory to approximately 75% of the node's available RAM. The remaining 25% accommodates the replication output buffer, AOF rewrite buffer, copy-on-write memory during RDB snapshots, and OS overhead. For a 16GB pod, set maxmemory to 12GB. For a 64GB bare metal server, set it to 48GB.

TLS Encryption and ACLs

Production Redis deployments must encrypt data in transit with TLS and enforce fine-grained access control with ACLs (Access Control Lists), introduced in Redis 6.0.

# redis.conf — TLS configuration
tls-port 6380
port 0                            # Disable non-TLS port entirely
tls-cert-file /etc/redis/tls/redis.crt
tls-key-file /etc/redis/tls/redis.key
tls-ca-cert-file /etc/redis/tls/ca.crt
tls-auth-clients optional         # Require client certificates (mutual TLS)
tls-replication yes               # Encrypt replication traffic
tls-cluster yes                   # Encrypt cluster bus traffic
tls-protocols "TLSv1.3"           # Only allow TLS 1.3

# ACL configuration (Redis 6.0+)
# user  on|off [>password] [~pattern] [+command|-command] [&channel]
user default off                  # Disable the default user
user admin on >strong_admin_pass ~* +@all
user appuser on >app_pass ~app:* ~session:* ~cache:* +@read +@write +@connection -@admin -@dangerous
user readonly on >readonly_pass ~* +@read +@connection -@write -@admin
user replicator on >repl_pass +psync +replconf +ping

# Load ACL from external file
aclfile /etc/redis/users.acl

Pub/Sub and Streams in HA Setups

Redis Pub/Sub and Streams behave differently in HA configurations. Understanding these differences is essential for building reliable event-driven systems.

Pub/Sub messages are fire-and-forget — they are not persisted, not replicated, and not buffered. In a Sentinel-based HA setup, subscribers connected to the master receive messages normally, but during a failover, the new master has no knowledge of previous subscriptions. Clients must re-subscribe after reconnecting. With Redis Cluster, Pub/Sub messages are broadcast to all nodes in the cluster, so subscribers connected to any node receive published messages (though this generates inter-node traffic).

Redis Streams are a persistent, replicated data structure that provides reliable message delivery in HA environments. Stream entries are replicated to replicas through the normal replication mechanism, survive failovers, and support consumer groups with at-least-once delivery semantics. For HA messaging, Streams should always be preferred over Pub/Sub.

# Redis Streams with consumer groups — HA-safe message processing

# Create a stream and consumer group
XGROUP CREATE events:orders orders-processors $ MKSTREAM

# Produce events
XADD events:orders * action "order_placed" order_id "12345" amount "599.99"
XADD events:orders * action "order_placed" order_id "12346" amount "149.99"

# Consume events (in consumer group — at-least-once delivery)
XREADGROUP GROUP orders-processors worker-1 COUNT 10 BLOCK 5000 STREAMS events:orders >

# Acknowledge processed events
XACK events:orders orders-processors 1681234567890-0

# Check pending messages (unacknowledged)
XPENDING events:orders orders-processors - + 10

# Claim abandoned messages (from a dead consumer)
XAUTOCLAIM events:orders orders-processors worker-2 60000 0-0 COUNT 10

# Trim stream to prevent unbounded growth
XTRIM events:orders MAXLEN ~ 100000

Redis Modules in HA

Redis modules extend Redis with specialised data structures and functionality. The three most popular — RedisJSON, RediSearch, and RedisTimeSeries — work with replication and Sentinel, but have specific considerations in HA deployments.

# redis.conf — Loading modules
loadmodule /opt/redis-stack/lib/rejson.so
loadmodule /opt/redis-stack/lib/redisearch.so
loadmodule /opt/redis-stack/lib/redistimeseries.so

# Modules are replicated to replicas via the command stream
# Ensure the same modules are installed on all nodes (master + replicas)

# RedisJSON — store and query JSON documents
JSON.SET user:1001 $ '{"name":"Alice","email":"alice@example.com","orders":42}'
JSON.GET user:1001 $.name

# RediSearch — full-text search with indexing
FT.CREATE idx:users ON JSON PREFIX 1 user: SCHEMA $.name AS name TEXT $.email AS email TAG $.orders AS orders NUMERIC
FT.SEARCH idx:users "@name:Alice"

# RedisTimeSeries — time-series data
TS.CREATE metrics:cpu:node1 RETENTION 86400000 LABELS host node1 metric cpu
TS.ADD metrics:cpu:node1 * 73.5
TS.RANGE metrics:cpu:node1 - + AGGREGATION avg 60000

When running Redis Stack (the module bundle) in HA, ensure all nodes — master and replicas — have identical module versions installed. Module commands are replicated via the standard replication stream, so replicas must be able to execute them. After a failover, RediSearch indexes exist on the promoted replica and serve queries immediately.

Connection Pooling and Client Configuration for HA

Proper connection pooling is essential for Redis HA performance and resilience. Connection pools reduce the overhead of establishing TCP connections, provide automatic retry logic, and enable graceful failover handling.

# Node.js — ioredis connection pool with cluster mode
const Redis = require('ioredis');

// Cluster mode connection
const cluster = new Redis.Cluster(
  [
    { host: '10.0.1.10', port: 7000 },
    { host: '10.0.1.11', port: 7000 },
    { host: '10.0.1.12', port: 7000 }
  ],
  {
    redisOptions: {
      password: 'cluster_password',
      connectTimeout: 10000,
      maxRetriesPerRequest: 3
    },
    scaleReads: 'slave',           // Route reads to replicas
    clusterRetryStrategy(times) {
      return Math.min(times * 200, 5000);
    },
    slotsRefreshTimeout: 2000,
    slotsRefreshInterval: 5000,
    enableOfflineQueue: true,
    enableReadyCheck: true,
    natMap: {}                     // For NAT/port-forwarded environments
  }
);

cluster.on('error', (err) => console.error('Cluster error:', err));
cluster.on('node error', (err, address) => {
  console.error(`Node ${address} error:`, err);
});

# Python — redis-py connection pool with cluster mode
from redis.cluster import RedisCluster
from redis.backoff import ExponentialBackoff
from redis.retry import Retry

retry = Retry(ExponentialBackoff(cap=5, base=0.1), retries=5)

rc = RedisCluster(
    startup_nodes=[
        {"host": "10.0.1.10", "port": 7000},
        {"host": "10.0.1.11", "port": 7000},
        {"host": "10.0.1.12", "port": 7000}
    ],
    password="cluster_password",
    decode_responses=True,
    read_from_replicas=True,
    retry=retry,
    retry_on_timeout=True,
    socket_timeout=5,
    socket_connect_timeout=5,
    max_connections=50,
    health_check_interval=30
)

rc.set("key", "value")
print(rc.get("key"))

// Go — go-redis cluster client with connection pooling
package main

import (
    "context"
    "time"

    "github.com/redis/go-redis/v9"
)

func NewRedisCluster() *redis.ClusterClient {
    return redis.NewClusterClient(&redis.ClusterOptions{
        Addrs: []string{
            "10.0.1.10:7000",
            "10.0.1.11:7000",
            "10.0.1.12:7000",
        },
        Password:         "cluster_password",
        ReadOnly:         true,
        RouteRandomly:    true,
        RouteByLatency:   false,
        PoolSize:         50,
        MinIdleConns:     10,
        DialTimeout:      10 * time.Second,
        ReadTimeout:      5 * time.Second,
        WriteTimeout:     5 * time.Second,
        PoolTimeout:      10 * time.Second,
        MaxRetries:       5,
        MinRetryBackoff:  200 * time.Millisecond,
        MaxRetryBackoff:  5 * time.Second,
    })
}

Backup and Restore Strategies

Even with HA replication, regular backups are essential for disaster recovery, compliance, and protection against logical errors (accidental FLUSHALL, bad application writes). Redis backups are based on RDB snapshots and AOF files.

#!/bin/bash
# redis-backup.sh — Automated Redis backup script

REDIS_HOST="10.0.1.10"
REDIS_PORT="6379"
REDIS_PASS="strong_master_password"
BACKUP_DIR="/backups/redis"
S3_BUCKET="s3://redis-backups-prod"
RETENTION_DAYS=30
DATE=$(date +%Y%m%d_%H%M%S)

mkdir -p "$BACKUP_DIR"

# Trigger RDB snapshot on the master
redis-cli -h "$REDIS_HOST" -p "$REDIS_PORT" -a "$REDIS_PASS" BGSAVE

# Wait for background save to complete
while [ "$(redis-cli -h $REDIS_HOST -p $REDIS_PORT -a $REDIS_PASS LASTSAVE)" = "$(redis-cli -h $REDIS_HOST -p $REDIS_PORT -a $REDIS_PASS LASTSAVE)" ]; do
  sleep 1
done
sleep 2

# Copy the RDB file
RDB_FILE=$(redis-cli -h "$REDIS_HOST" -p "$REDIS_PORT" -a "$REDIS_PASS" CONFIG GET dir | tail -1)
cp "${RDB_FILE}/dump.rdb" "${BACKUP_DIR}/dump_${DATE}.rdb"

# Compress and upload to S3
gzip "${BACKUP_DIR}/dump_${DATE}.rdb"
aws s3 cp "${BACKUP_DIR}/dump_${DATE}.rdb.gz" "${S3_BUCKET}/daily/dump_${DATE}.rdb.gz" \
  --storage-class STANDARD_IA

# Cleanup old local backups
find "$BACKUP_DIR" -name "dump_*.rdb.gz" -mtime +$RETENTION_DAYS -delete

# Verify backup integrity
redis-check-rdb "${BACKUP_DIR}/dump_${DATE}.rdb.gz" && \
  echo "Backup verified: dump_${DATE}.rdb.gz" || \
  echo "ERROR: Backup verification failed!"

echo "Backup complete: ${BACKUP_DIR}/dump_${DATE}.rdb.gz"

# Restore from RDB backup
# 1. Stop Redis
systemctl stop redis

# 2. Replace the RDB file
gunzip /backups/redis/dump_20260412_030000.rdb.gz
cp /backups/redis/dump_20260412_030000.rdb /data/redis/dump.rdb
chown redis:redis /data/redis/dump.rdb

# 3. Disable AOF temporarily (if enabled) to prevent AOF overriding RDB on startup
redis-cli -a password CONFIG SET appendonly no

# 4. Start Redis (loads RDB)
systemctl start redis

# 5. Re-enable AOF and rewrite it from the loaded data
redis-cli -a password CONFIG SET appendonly yes
redis-cli -a password BGREWRITEAOF

Monitoring with Redis INFO, Prometheus, and Grafana

Comprehensive monitoring is the foundation of operational excellence for Redis HA. Redis exposes rich internal metrics through the INFO command, which the Prometheus Redis exporter translates into time-series metrics for Grafana dashboards and alerting.

# Key Redis INFO sections for HA monitoring
redis-cli -a password INFO replication
# role:master
# connected_slaves:2
# slave0:ip=10.0.1.11,port=6379,state=online,offset=1234567,lag=0
# slave1:ip=10.0.1.12,port=6379,state=online,offset=1234560,lag=1
# master_replid:abc123...
# master_repl_offset:1234567
# repl_backlog_size:268435456
# repl_backlog_first_byte_offset:1000000

redis-cli -a password INFO memory
# used_memory:6442450944
# used_memory_human:6.00G
# used_memory_rss:6879707136
# mem_fragmentation_ratio:1.07
# maxmemory:6442450944
# maxmemory_policy:allkeys-lru
# evicted_keys:12345

redis-cli -a password INFO stats
# total_connections_received:50000
# total_commands_processed:12345678
# instantaneous_ops_per_sec:85432
# keyspace_hits:11000000
# keyspace_misses:1345678
# expired_keys:500000
# evicted_keys:12345

redis-cli -a password INFO clients
# connected_clients:150
# blocked_clients:0
# maxclients:10000

# Prometheus Redis Exporter deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: redis-exporter
  namespace: monitoring
spec:
  replicas: 1
  selector:
    matchLabels:
      app: redis-exporter
  template:
    metadata:
      labels:
        app: redis-exporter
      annotations:
        prometheus.io/scrape: "true"
        prometheus.io/port: "9121"
    spec:
      containers:
        - name: redis-exporter
          image: oliver006/redis_exporter:latest
          args:
            - --redis.addr=redis://redis-ha-master:6379
            - --redis.password=$(REDIS_PASSWORD)
            - --include-system-metrics
            - --is-cluster
          env:
            - name: REDIS_PASSWORD
              valueFrom:
                secretKeyRef:
                  name: redis-secret
                  key: password
          ports:
            - containerPort: 9121
              name: metrics
          resources:
            requests:
              cpu: 100m
              memory: 128Mi
            limits:
              cpu: 200m
              memory: 256Mi

# PrometheusRule for Redis HA alerts
apiVersion: monitoring.coreos.com/v1
kind: PrometheusRule
metadata:
  name: redis-ha-alerts
  namespace: monitoring
spec:
  groups:
    - name: redis-availability
      rules:
        - alert: RedisDown
          expr: redis_up == 0
          for: 1m
          labels:
            severity: critical
          annotations:
            summary: "Redis instance {{ $labels.instance }} is down"

        - alert: RedisReplicaDisconnected
          expr: redis_connected_slaves < 2
          for: 2m
          labels:
            severity: warning
          annotations:
            summary: "Redis master has fewer than 2 connected replicas"

        - alert: RedisReplicationLagHigh
          expr: redis_replication_lag > 5
          for: 3m
          labels:
            severity: warning
          annotations:
            summary: "Redis replication lag exceeds 5 seconds on {{ $labels.instance }}"

        - alert: RedisMemoryUsageHigh
          expr: redis_memory_used_bytes / redis_memory_max_bytes > 0.9
          for: 5m
          labels:
            severity: warning
          annotations:
            summary: "Redis memory usage above 90% on {{ $labels.instance }}"

        - alert: RedisEvictionsHigh
          expr: rate(redis_evicted_keys_total[5m]) > 100
          for: 5m
          labels:
            severity: warning
          annotations:
            summary: "Redis evicting keys at >100/s on {{ $labels.instance }}"

        - alert: RedisKeyspaceHitRateLow
          expr: redis_keyspace_hits_total / (redis_keyspace_hits_total + redis_keyspace_misses_total) < 0.8
          for: 10m
          labels:
            severity: info
          annotations:
            summary: "Redis cache hit rate below 80% on {{ $labels.instance }}"

        - alert: RedisSentinelDown
          expr: redis_sentinel_master_status != 1
          for: 1m
          labels:
            severity: critical
          annotations:
            summary: "Redis Sentinel reports master unhealthy"

    - name: redis-performance
      rules:
        - alert: RedisSlowlogGrowing
          expr: increase(redis_slowlog_length[5m]) > 10
          for: 5m
          labels:
            severity: warning
          annotations:
            summary: "Redis slow log growing rapidly on {{ $labels.instance }}"

        - alert: RedisConnectionsNearLimit
          expr: redis_connected_clients / redis_config_maxclients > 0.8
          for: 5m
          labels:
            severity: warning
          annotations:
            summary: "Redis connected clients above 80% of maxclients"

Dragonfly and KeyDB: Redis-Compatible Alternatives

While Redis is the dominant in-memory data store, two Redis-compatible alternatives have gained traction for specific use cases.

Dragonfly

Dragonfly is a modern, multi-threaded Redis replacement that aims to be a drop-in replacement while leveraging all available CPU cores. Traditional Redis is single-threaded for command processing — Dragonfly uses a shared-nothing architecture with multiple threads to achieve significantly higher throughput on multi-core machines. It supports the Redis protocol, most Redis commands, and can replace Redis without application changes.

# Run Dragonfly as a Redis replacement
docker run -d --name dragonfly \
  -p 6379:6379 \
  -v /data/dragonfly:/data \
  docker.dragonflydb.io/dragonflydb/dragonfly \
  --maxmemory 12gb \
  --proactor_threads 8 \
  --dbfilename dump.rdb \
  --requirepass strong_password

# Dragonfly in Kubernetes
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: dragonfly
  namespace: production
spec:
  replicas: 1
  selector:
    matchLabels:
      app: dragonfly
  template:
    metadata:
      labels:
        app: dragonfly
    spec:
      containers:
        - name: dragonfly
          image: docker.dragonflydb.io/dragonflydb/dragonfly:latest
          args:
            - --maxmemory=12gb
            - --proactor_threads=8
            - --requirepass=strong_password
            - --snapshot_cron=*/30 * * * *
          ports:
            - containerPort: 6379
          resources:
            requests:
              cpu: "4"
              memory: 16Gi
            limits:
              cpu: "8"
              memory: 16Gi
          volumeMounts:
            - name: data
              mountPath: /data
  volumeClaimTemplates:
    - metadata:
        name: data
      spec:
        accessModes: ["ReadWriteOnce"]
        storageClassName: longhorn
        resources:
          requests:
            storage: 100Gi

Dragonfly's key advantages: multi-threaded (25x throughput on 8 cores vs single-threaded Redis), better memory efficiency (uses Dash hash tables instead of Redis dict), built-in snapshotting without fork() overhead, and native support for large datasets. However, as of 2026, Dragonfly's replication support is still maturing — it supports primary-replica replication but does not yet have a Sentinel-equivalent automatic failover system. For HA, use Kubernetes health checks and StatefulSet restart policies, or deploy behind a load balancer with application-level failover.

KeyDB

KeyDB is a multi-threaded fork of Redis maintained by Snap (the company behind Snapchat). It is fully compatible with Redis and adds multi-threading, active-active replication (multi-master), FLASH storage tiering, and subkey expiration. KeyDB's active-active replication is particularly interesting for HA — two KeyDB instances can accept writes simultaneously and replicate to each other, providing zero-downtime failover.

# keydb.conf — Multi-threaded configuration with active replication
server-threads 4                  # Use 4 threads for command processing
bind 0.0.0.0
port 6379
requirepass strong_password
masterauth strong_password

# Active-active replication (multi-master)
active-replica yes
replicaof peer-host 6379          # Bidirectional replication

# On the peer node, configure the reverse:
# replicaof this-host 6379

# FLASH storage tiering (for datasets larger than RAM)
# storage-provider flash /mnt/flash-storage 100
# maxmemory 16gb
# Will keep hot data in RAM and spill cold data to SSD

# SubKey expiration (unique to KeyDB)
# Allows setting TTL on hash fields, not just top-level keys
# EXPIREMEMBER myhash field1 3600

KeyDB is a strong choice when you need multi-master replication for geographic distribution or zero-downtime maintenance, or when you need higher throughput than single-threaded Redis can provide but want to stay closer to the Redis codebase than Dragonfly.

Performance Tuning

Pipelining

Pipelining batches multiple commands into a single network round-trip, dramatically reducing latency for bulk operations. Instead of waiting for each response before sending the next command, the client sends all commands at once and reads all responses together.

# Python — Pipelining with redis-py
import redis
import time

r = redis.Redis(host='10.0.1.10', port=6379, password='password', decode_responses=True)

# Without pipelining: 1000 round-trips
start = time.time()
for i in range(1000):
    r.set(f'key:{i}', f'value:{i}')
print(f'Without pipeline: {time.time() - start:.3f}s')

# With pipelining: 1 round-trip for 1000 commands
start = time.time()
pipe = r.pipeline(transaction=False)
for i in range(1000):
    pipe.set(f'key:{i}', f'value:{i}')
pipe.execute()
print(f'With pipeline: {time.time() - start:.3f}s')
# Typically 5-10x faster

Lua Scripting

Lua scripts execute atomically on the Redis server, eliminating round-trips for complex operations and guaranteeing that no other command executes between the script's operations. In Redis Cluster, ensure all keys accessed by a Lua script reside in the same hash slot using hash tags.

# Lua script for atomic rate limiting
# KEYS[1] = rate limit key
# ARGV[1] = max requests
# ARGV[2] = window in seconds
local current = redis.call('INCR', KEYS[1])
if current == 1 then
    redis.call('EXPIRE', KEYS[1], ARGV[2])
end
if current > tonumber(ARGV[1]) then
    return 0  -- Rate limited
end
return 1  -- Allowed

# Load and execute
redis-cli -a password EVAL "\
  local current = redis.call('INCR', KEYS[1]) \
  if current == 1 then redis.call('EXPIRE', KEYS[1], ARGV[2]) end \
  if current > tonumber(ARGV[1]) then return 0 end \
  return 1" 1 ratelimit:user:123 100 60

Memory Optimisation

# redis.conf — Memory optimisation settings

# Use ziplist encoding for small hashes, lists, sorted sets
hash-max-listpack-entries 128
hash-max-listpack-value 64
list-max-listpack-size -2          # 8KB per node
zset-max-listpack-entries 128
zset-max-listpack-value 64
set-max-intset-entries 512

# Lazy freeing (avoid blocking on large key deletion)
lazyfree-lazy-eviction yes
lazyfree-lazy-expire yes
lazyfree-lazy-server-del yes
lazyfree-lazy-user-del yes
lazyfree-lazy-user-flush yes

# jemalloc tuning
# MALLOC_CONF="background_thread:true,dirty_decay_ms:5000,muzzy_decay_ms:5000"

# Analyse memory usage
redis-cli -a password MEMORY DOCTOR
redis-cli -a password MEMORY STATS
redis-cli -a password --bigkeys
redis-cli -a password --memkeys

Common Failure Scenarios and Troubleshooting

Scenario 1: Master Fails, Sentinel Promotes Replica

# Diagnosis
redis-cli -p 26379 SENTINEL master mymaster
# Check: flags should show 's_down' or 'o_down' if master is unreachable

redis-cli -p 26379 SENTINEL get-master-addr-by-name mymaster
# Returns the current master address (should be the promoted replica)

# Check Sentinel logs for failover events
tail -f /var/log/redis/sentinel.log
# Look for: +sdown, +odown, +try-failover, +elected-leader,
#           +failover-state-select-slave, +selected-slave,
#           +failover-state-send-slaveof-noone, +failover-end

Scenario 2: Split-Brain with Two Masters

# Diagnosis: check if min-replicas-to-write is configured
redis-cli -a password CONFIG GET min-replicas-to-write
redis-cli -a password CONFIG GET min-replicas-max-lag

# Prevention: configure min-replicas on all masters
redis-cli -a password CONFIG SET min-replicas-to-write 1
redis-cli -a password CONFIG SET min-replicas-max-lag 10

# If split-brain occurred: identify the stale master
# Compare replication offsets — the master with the higher offset has more data
redis-cli -h master1-ip -a password INFO replication | grep master_repl_offset
redis-cli -h master2-ip -a password INFO replication | grep master_repl_offset

# Force the stale master to become a replica
redis-cli -h stale-master-ip -a password REPLICAOF correct-master-ip 6379

Scenario 3: Full Resync Storm After Network Partition

# Diagnosis: check replication backlog
redis-cli -a password INFO replication | grep repl_backlog
# If repl_backlog_first_byte_offset is ahead of replica's offset, full sync triggers

# Prevention: increase backlog size
redis-cli -a password CONFIG SET repl-backlog-size 512mb

# Monitor for full syncs
redis-cli -a password INFO stats | grep sync_full
redis-cli -a password INFO stats | grep sync_partial_ok
redis-cli -a password INFO stats | grep sync_partial_err

Scenario 4: Memory Exhaustion and OOM Kill

# Diagnosis
redis-cli -a password INFO memory
# Check: used_memory vs maxmemory, mem_fragmentation_ratio

redis-cli -a password MEMORY DOCTOR
# Returns advice on memory issues

# Prevention: set proper maxmemory and eviction
redis-cli -a password CONFIG SET maxmemory 12gb
redis-cli -a password CONFIG SET maxmemory-policy allkeys-lru

# Find large keys consuming memory
redis-cli -a password --bigkeys
redis-cli -a password --memkeys --memkeys-samples 100

# Emergency: manually evict keys
redis-cli -a password SCAN 0 COUNT 1000 TYPE string
# Identify and DEL unnecessary large keys

Scenario 5: Slow Commands Blocking Replication

# Diagnosis: check slowlog
redis-cli -a password SLOWLOG GET 20
redis-cli -a password SLOWLOG LEN

# Check for blocking commands
redis-cli -a password CLIENT LIST | grep -E 'cmd=(keys|sort|smembers)'

# Prevention: configure slowlog threshold
redis-cli -a password CONFIG SET slowlog-log-slower-than 10000   # 10ms
redis-cli -a password CONFIG SET slowlog-max-len 256

# Rename dangerous commands
rename-command KEYS ""              # Disable KEYS entirely
rename-command FLUSHALL ""          # Disable FLUSHALL
rename-command FLUSHDB ""           # Disable FLUSHDB
rename-command DEBUG ""             # Disable DEBUG

Capacity Planning and Scaling Strategies

Capacity planning for Redis HA involves estimating memory requirements, network bandwidth, and CPU utilisation across master and replica nodes. The key metrics to plan around are dataset size, operations per second, average key/value size, and replication overhead.

# Capacity estimation formulas

# Memory per node:
#   Base dataset size (use redis-cli DBSIZE and MEMORY USAGE on a sample)
#   + Replication output buffer: ~64MB per replica
#   + AOF rewrite buffer: ~64MB during rewrites
#   + Copy-on-write overhead during BGSAVE: up to 2x during heavy writes
#   + Client output buffers: ~1KB per client
#   + OS overhead: ~1-2GB
#   Rule of thumb: maxmemory = 75% of available RAM

# Network bandwidth:
#   Replication: write_throughput_bytes * num_replicas
#   Client traffic: ops_per_sec * avg_response_size
#   Full sync: dataset_size (one-time during replica bootstrap or failover)

# Example sizing for 20GB dataset, 100K ops/sec:
#   RAM per node: 20GB data + 4GB buffers + 2GB OS = 26GB -> 32GB node (75% = 24GB maxmemory)
#   CPU: 1 core handles ~100K ops/sec for simple commands (GET/SET)
#   Network: 100K ops * 1KB avg = 100MB/s client + 50MB/s replication = 150MB/s per master

# Scaling decision tree:
#   Need more read throughput? -> Add replicas (up to 5 per master)
#   Need more write throughput? -> Redis Cluster (add shards)
#   Need more memory? -> Redis Cluster (distribute dataset across shards)
#   Need lower latency? -> Reduce network hops (co-locate, use unix sockets)
#   Need global distribution? -> Multi-region replication or Redis Enterprise Active-Active

Horizontal Scaling with Redis Cluster

# Add shards to an existing Redis Cluster

# 1. Start new Redis nodes
redis-server /etc/redis/new-master.conf
redis-server /etc/redis/new-replica.conf

# 2. Add the new master to the cluster
redis-cli --cluster add-node new-master:7000 existing-node:7000 -a password

# 3. Add the new replica to follow the new master
redis-cli --cluster add-node new-replica:7000 existing-node:7000 \
  --cluster-slave --cluster-master-id NEW_MASTER_ID -a password

# 4. Reshard slots to the new master
redis-cli --cluster reshard existing-node:7000 \
  --cluster-from all --cluster-to NEW_MASTER_ID \
  --cluster-slots 4096 --cluster-yes -a password

# 5. Verify the new slot distribution
redis-cli -c -h existing-node -p 7000 -a password CLUSTER SLOTS

# Remove a shard (scale down)
# 1. Reshard all slots away from the node
redis-cli --cluster reshard existing-node:7000 \
  --cluster-from REMOVING_NODE_ID --cluster-to TARGET_NODE_ID \
  --cluster-slots 5461 --cluster-yes -a password

# 2. Remove the empty node
redis-cli --cluster del-node existing-node:7000 REMOVING_NODE_ID -a password

Vertical Scaling Considerations

# When to scale vertically vs horizontally:

# Scale UP (bigger instances) when:
#   - Dataset fits in single-node memory
#   - Workload uses multi-key operations (MGET, SUNION, Lua across keys)
#   - Operational simplicity is more important than cost efficiency
#   - Using Redis modules that don't support Cluster mode well

# Scale OUT (more shards) when:
#   - Dataset exceeds single-node memory
#   - Write throughput exceeds single-thread capacity (~200K ops/sec)
#   - You need per-shard isolation for multi-tenant workloads
#   - Cost per GB of RAM is a concern (many smaller nodes vs few large ones)

# Cloud instance recommendations:
# AWS:   cache.r7g.xlarge (4 vCPU, 26GB) to cache.r7g.16xlarge (64 vCPU, 419GB)
# Azure: P1 (6GB) to P5 (120GB) per shard
# GCP:   5GB to 300GB per instance (Memorystore Standard)

# Bare metal:
#   CPU: 2-4 cores dedicated to Redis (single-threaded, but background tasks use extra cores)
#   RAM: 32-128GB per node (NVMe for swap-as-last-resort)
#   Network: 10Gbps minimum, 25Gbps for large datasets
#   Storage: NVMe SSD for AOF/RDB persistence (IOPS matters for fsync)

Conclusion

Redis high availability is not a single configuration choice — it is a comprehensive system design that spans replication topology, failure detection, automatic failover, client configuration, persistence strategy, memory management, security, monitoring, and operational procedures. The right HA architecture depends on your specific requirements.

For most applications, Redis Sentinel with three Sentinel instances monitoring a master and two replicas provides a proven, battle-tested HA solution with automatic failover in under 30 seconds. When you need horizontal scaling beyond a single master's throughput or memory capacity, Redis Cluster distributes the dataset across multiple shards while maintaining built-in failover per shard. For Kubernetes environments, operators like Spotahome and OpsTree encode operational best practices into declarative CRDs, while Redis Enterprise provides the most feature-rich option with Active-Active geo-replication and module support.

Cloud-managed services — AWS ElastiCache, Azure Cache for Redis, and GCP Memorystore — eliminate the operational burden of running Redis infrastructure but come with reduced flexibility and higher costs at scale. For organisations with bare metal infrastructure, k3s with Rancher and Longhorn provides a fully open-source, cloud-independent alternative that delivers enterprise-grade HA without vendor lock-in.

Alternatives like Dragonfly and KeyDB are worth evaluating for specific use cases — Dragonfly for raw multi-threaded throughput on large machines, and KeyDB for active-active multi-master replication. Both are Redis-compatible and can serve as drop-in replacements in many scenarios.

Regardless of which architecture you choose, the operational fundamentals remain constant: configure proper persistence (hybrid RDB + AOF), enforce min-replicas-to-write to prevent split-brain data loss, size your replication backlog for your write volume, encrypt all traffic with TLS, enforce least-privilege access with ACLs, monitor replication lag and memory usage with Prometheus and Grafana, back up RDB snapshots to durable off-site storage, and — most critically — test your failover regularly. A failover system that has never been tested is a system that does not work. Run monthly failover drills, inject failures with chaos engineering tools, and measure your actual recovery time. The confidence you gain from systematic testing is what separates a Redis deployment that survives production incidents from one that turns a server failure into a company-wide outage.