Redis

是什么

Redis 是内存数据结构服务器，提供缓存、计数、排行、消息、分布式锁等一站式高速能力。 用它的效果是：把数据库扛不住的高频读写卸载到内存层，让系统在峰值依然稳得住。

怎么用

1. 先按场景选择数据结构（String、Hash、Sorted Set、Stream），让性能与表达力都不浪费。
2. 为每个键设置过期策略（TTL），让缓存不会无限增长拖垮内存。
3. 用 Lua 脚本或事务封装多步操作，让并发场景下的一致性有保证。
4. 通过哨兵（Sentinel）或集群（Cluster）方案规划高可用与扩展。
5. 持续监控命中率与内存使用，让缓存策略可以基于真实数据而非直觉调整。

架构图

flowchart LR
  应用 --> 缓存读取
  缓存读取 --> 命中返回
  缓存读取 --> 回源数据库
  回源数据库 --> 写回缓存
  写回缓存 --> 过期淘汰

Redis Patterns

Redis data structures, caching, pub/sub, and streams for Python async apps.

When to Activate

• Choosing the right Redis data structure for a use case
• Implementing caching (cache-aside, write-through, TTL eviction)
• Using pub/sub or Redis Streams for real-time messaging / SSE
• Building a job queue with SKIP LOCKED semantics
• Writing atomic operations (counters, rate limits, distributed locks)
• Debugging slow Redis commands or memory bloat

---

Connection (async redis-py)

import redis.asyncio as redis

# Single connection
client = await redis.from_url("redis://localhost:6379", decode_responses=True)

# Connection pool (recommended for apps)
pool = redis.ConnectionPool.from_url(
    "redis://localhost:6379",
    decode_responses=True,
    max_connections=20,
)
client = redis.Redis(connection_pool=pool)

# Close on shutdown
await client.aclose()

decode_responses=True returns str instead of bytes — use it unless you store binary data.

---

Data Structures

Strings — single values, counters, JSON blobs

# Set / get
await client.set("user:123:name", "Alice")
await client.get("user:123:name")            # "Alice"

# With TTL (seconds)
await client.set("session:abc", token, ex=3600)    # expires in 1 hour
await client.setex("session:abc", 3600, token)     # same

# Only set if not exists (NX) — distributed lock primitive
acquired = await client.set("lock:job:42", "worker-1", nx=True, ex=30)

# Atomic counter
await client.incr("page:views")
await client.incrby("page:views", 5)
await client.decr("inventory:product:99")

# Get + set atomically (Lua or GETEX)
await client.getex("session:abc", ex=3600)   # reset TTL on read

# Store JSON
import json
await client.set("user:123", json.dumps(user_dict))
user = json.loads(await client.get("user:123"))

Hashes — objects / partial updates

# Set multiple fields at once
await client.hset("user:123", mapping={
    "name": "Alice",
    "email": "[email protected]",
    "role": "admin",
})

# Get all fields
user = await client.hgetall("user:123")     # {"name": "Alice", ...}

# Get one field
name = await client.hget("user:123", "name")

# Update one field without overwriting others
await client.hset("user:123", "role", "user")

# Check existence
exists = await client.hexists("user:123", "email")

# Delete a field
await client.hdel("user:123", "temp_token")

# Get field names / values
fields = await client.hkeys("user:123")
values = await client.hvals("user:123")

Use hashes for objects with many fields where you update individual fields often. Cheaper than JSON string for partial reads.

Lists — queues, activity feeds

# Push to right (tail) — enqueue
await client.rpush("queue:emails", "msg-1", "msg-2")

# Pop from left (head) — dequeue FIFO
job = await client.lpop("queue:emails")

# Blocking pop — wait up to 30s for an item
job = await client.blpop("queue:emails", timeout=30)   # returns (key, value)

# Stack (LIFO): rpush + rpop
await client.rpush("stack", "item")
item = await client.rpop("stack")

# Peek without removing
items = await client.lrange("queue:emails", 0, -1)   # all items
recent = await client.lrange("activity:user:1", 0, 9)  # first 10

# Keep list bounded (trim to last 100)
await client.ltrim("activity:user:1", -100, -1)

# Length
length = await client.llen("queue:emails")

Sets — unique membership, tags

await client.sadd("online_users", "user-1", "user-2")
await client.srem("online_users", "user-2")

is_online = await client.sismember("online_users", "user-1")
members = await client.smembers("online_users")
count = await client.scard("online_users")

# Set operations
common = await client.sinter("user:1:friends", "user:2:friends")  # intersection
all_  = await client.sunion("tag:python", "tag:async")             # union
diff  = await client.sdiff("all_users", "banned_users")            # difference

Sorted Sets — leaderboards, priority queues, rate limiting

# Add with score (score determines order)
await client.zadd("leaderboard", {"alice": 1500, "bob": 1200, "carol": 1800})

# Get top 3 (highest score first)
top3 = await client.zrevrange("leaderboard", 0, 2, withscores=True)
# [("carol", 1800.0), ("alice", 1500.0), ("bob", 1200.0)]

# Rank (0-indexed, lowest score = rank 0)
rank = await client.zrevrank("leaderboard", "alice")   # 1 (2nd place)

# Increment score atomically
await client.zincrby("leaderboard", 50, "bob")

# Range by score — get items between two scores
members = await client.zrangebyscore("leaderboard", 1400, 2000)

# Remove
await client.zrem("leaderboard", "bob")

---

TTL and Expiration

# Set TTL on existing key
await client.expire("session:abc", 3600)           # seconds
await client.expireat("session:abc", timestamp)    # unix timestamp
await client.pexpire("key", 500)                   # milliseconds

# Check remaining TTL
ttl = await client.ttl("session:abc")    # seconds remaining, -1 if no TTL, -2 if missing
pttl = await client.pttl("session:abc") # milliseconds

# Remove TTL (make persistent)
await client.persist("key")

---

Pub/Sub

# Publisher
async def publish_event(client, channel: str, data: dict):
    await client.publish(channel, json.dumps(data))

# Subscriber — runs indefinitely
async def subscribe_to_events(client, channel: str):
    async with client.pubsub() as pubsub:
        await pubsub.subscribe(channel)
        async for message in pubsub.listen():
            if message["type"] == "message":
                data = json.loads(message["data"])
                yield data

# Pattern subscribe
async with client.pubsub() as pubsub:
    await pubsub.psubscribe("tasks:*")   # matches tasks:created, tasks:done, etc.
    async for message in pubsub.listen():
        if message["type"] == "pmessage":
            process(message["channel"], message["data"])

Limitation: pub/sub messages are fire-and-forget. Subscribers that miss a message while offline don't receive it. Use Streams for durable delivery.

---

Redis Streams (durable pub/sub)

Streams persist messages — consumers can read from any position, including past messages.

# Produce — append message to stream
msg_id = await client.xadd(
    "task:updates",
    {"task_id": "t-123", "status": "running", "content": "Processing..."},
    maxlen=10000,     # trim to 10k entries (approximate)
)

# Consume from beginning
messages = await client.xread({"task:updates": "0-0"}, count=100)
# messages: [("task:updates", [(id, {fields...}), ...])]

# Consume only new messages (since last read)
last_id = "0-0"
messages = await client.xread({"task:updates": last_id}, count=10, block=5000)
for stream, entries in messages:
    for msg_id, fields in entries:
        process(fields)
        last_id = msg_id

# Consumer groups — multiple workers compete for messages
await client.xgroup_create("task:updates", "workers", id="0", mkstream=True)

# Worker reads and claims a message
msgs = await client.xreadgroup("workers", "worker-1", {"task:updates": ">"}, count=1)
for stream, entries in msgs:
    for msg_id, fields in entries:
        process(fields)
        await client.xack("task:updates", "workers", msg_id)   # mark done

# Trim old entries
await client.xtrim("task:updates", maxlen=5000, approximate=True)

SSE streaming pattern (used in Agentex frontend):

# Backend: push deltas to a stream per task
await client.xadd(f"task:{task_id}:stream", {"delta": chunk})

# Frontend SSE endpoint: read stream and forward to browser
async def stream_task(task_id: str):
    last_id = "0-0"
    while True:
        messages = await client.xread({f"task:{task_id}:stream": last_id}, block=5000)
        for _, entries in messages:
            for msg_id, fields in entries:
                yield f"data: {fields['delta']}\n\n"
                last_id = msg_id

---

Caching Patterns

Cache-aside (lazy loading)

async def get_user(user_id: str) -> User:
    key = f"user:{user_id}"
    cached = await client.get(key)
    if cached:
        return User(**json.loads(cached))

    user = await db.fetch_user(user_id)
    await client.set(key, user.model_dump_json(), ex=300)   # cache 5 min
    return user

async def invalidate_user(user_id: str):
    await client.delete(f"user:{user_id}")

Write-through

async def update_user(user_id: str, data: dict) -> User:
    user = await db.update_user(user_id, data)
    await client.set(f"user:{user_id}", user.model_dump_json(), ex=300)
    return user

---

Atomic Operations

Distributed lock

import uuid

async def with_lock(client, resource: str, ttl: int = 30):
    lock_key = f"lock:{resource}"
    lock_val = str(uuid.uuid4())

    acquired = await client.set(lock_key, lock_val, nx=True, ex=ttl)
    if not acquired:
        raise RuntimeError(f"Could not acquire lock on {resource}")
    try:
        yield
    finally:
        # Only release if we still own it (Lua script for atomicity)
        script = """
        if redis.call("get", KEYS[1]) == ARGV[1] then
            return redis.call("del", KEYS[1])
        else
            return 0
        end
        """
        await client.eval(script, 1, lock_key, lock_val)

Rate limiting (sliding window)

async def is_rate_limited(client, user_id: str, limit: int = 100, window: int = 60) -> bool:
    key = f"rate:{user_id}:{int(time.time()) // window}"
    count = await client.incr(key)
    if count == 1:
        await client.expire(key, window)
    return count > limit

Pipeline (batch commands — reduce round trips)

async with client.pipeline(transaction=False) as pipe:
    pipe.hset("user:1", mapping=data)
    pipe.expire("user:1", 3600)
    pipe.zadd("leaderboard", {"user-1": score})
    results = await pipe.execute()   # sent as one network round trip

# Atomic pipeline (MULTI/EXEC)
async with client.pipeline(transaction=True) as pipe:
    await pipe.watch("inventory:42")
    quantity = int(await pipe.get("inventory:42"))
    if quantity < 1:
        raise Exception("Out of stock")
    pipe.multi()
    pipe.decr("inventory:42")
    await pipe.execute()

---

Red Flags

• No TTL on cache or session keys — keys without expiry accumulate forever and evict randomly under memory pressure; set ex= on every set() call for cached data and sessions
• Using pub/sub for reliable delivery — pub/sub is fire-and-forget; subscribers that are offline when a message is published never receive it; use Redis Streams with consumer groups for any message that must not be lost
• Single connection instead of a pool — a single await redis.from_url(...) connection serializes all commands and blocks under concurrent load; use ConnectionPool with max_connections sized to your concurrency
• KEYS * in production — KEYS is O(n) and blocks the Redis event loop while it scans every key; use SCAN with a cursor to iterate non-blocking, or redesign to avoid key enumeration entirely
• Distributed lock without a unique value — a lock released by any caller using only the key (not the unique lock value) can accidentally release another owner's lock; always store a UUID as the value and use a Lua script to compare-then-delete atomically
• Unbounded stream growth — xadd without maxlen lets the stream grow indefinitely; always set maxlen=N (with approximate=True for efficiency) or run periodic xtrim
• Sending multiple independent commands one at a time — each await client.set(...) is a network round trip; batch three or more independent commands in a pipeline (async with client.pipeline()) to cut round-trip overhead significantly

Checklist

• Connection pool used (not single connection) for async apps
• decode_responses=True set unless storing binary
• TTL set on all cache/session keys
• Pub/sub replaced with Streams where offline delivery matters
• xack called after processing stream messages (consumer groups)
• Pipeline used when sending ≥ 3 independent commands in sequence
• Distributed locks use NX + expiry to prevent deadlocks
• Sorted sets used for leaderboards / time-ordered data instead of sorted lists
• maxlen set on streams to prevent unbounded growth