Module P-9·16 min read

Storing sessions as Hashes vs JSON strings, sliding expiry with EXPIRE on each request, session invalidation and logout, multi-device session tracking, and the consistency trade-offs when reading sessions from replicas.

P-9 — Session Management Patterns

Q: A team decides to store user sessions as JSON strings in Redis (`SET session:123 '{"userId":1, "lastSeen":170000}'`). They have an endpoint that updates the `lastSeen` timestamp and another concurrent endpoint that adds a new item to a `cart` array within the same session. What is the fundamental risk of this approach compared to using a Redis Hash?

JSON strings require a read-modify-write cycle in the application. Under concurrency, this creates a race condition where one endpoint's update might overwrite and destroy the other endpoint's changes. — This is a classic concurrency race condition. If two requests fetch the same JSON string simultaneously, modify different parts of it in application memory, and then both save it back, the last one to save will overwrite the first one's changes. Redis Hashes solve this by allowing atomic, field-level updates (e.g., `HSET session:123 lastSeen 170000`), meaning the `lastSeen` update doesn't touch or overwrite the `cart` data.

Q: An application enforces both an "idle timeout" (30 minutes of inactivity) and an "absolute timeout" (24 hours maximum session life). Which implementation strategy is correct and most efficient?

Store `absoluteExpiry` inside the session Hash. Set the Redis key TTL to 1800 (30 minutes) and reset this TTL on every user interaction. On every read, compare `Date.now()` to the stored `absoluteExpiry` and delete the session if it has passed. — Redis's native TTL mechanism only supports one expiration time per key. The most robust way to handle dual timeouts is to map the volatile "idle timeout" to the Redis key's actual TTL (resetting it to 30 minutes on every touch) and embed the hard "absolute timeout" as a data field within the session itself. This ensures Redis handles the sliding expiry automatically, while the application strictly enforces the hard cutoff upon reading the session data.

Q: To improve read performance, an engineering team configures their Node.js application to read session data from Redis read-replicas, while continuing to write new sessions to the primary. Shortly after, users report occasionally being "logged out" immediately after successfully logging in. What is the cause?

Redis replication is asynchronous. The user logs in, the session is written to the primary, and the redirect sends them to the dashboard. The dashboard read hits a replica *before* the session data has replicated across the network, resulting in a cache miss and a false logout. — This is the classic "Read-from-Replica Consistency Trap". Redis replication is highly performant but inherently asynchronous (eventually consistent). For critical, immediate-read-after-write operations like session authentication, the small replication delay (even if just a few milliseconds) creates a race condition. The solution is simple: always read session data from the primary node.

Who this module is for: You store user sessions in Redis — or are planning to — and want to implement them correctly. This covers session structure (Hash vs JSON string), sliding expiry, concurrent device management, and the read-from-replica consistency trap that silently logs users out.

Why Redis for Sessions

HTTP is stateless. Sessions provide continuity: a session token in a cookie maps to server-side state (user ID, permissions, preferences). The session store must be:

Fast — checked on every request
Shared — accessible from every application server instance
Expirable — sessions must auto-expire when idle or after a fixed duration

Redis satisfies all three. A 1ms session lookup is imperceptible. Every application server connects to the same Redis instance. TTL handles expiry automatically.

The alternative — database-backed sessions (PostgreSQL, MySQL) — works but adds a database query to every request, and relational databases are optimized for complex queries, not millions of simple ID lookups.

Session Structure: Hash vs JSON String

Option A: JSON String

SET session:{token} '{"userId":1001,"email":"j@example.com","role":"engineer","permissions":["read","write"]}' EX 3600

Simple. One key, one value. But updating a single field requires:

GET session:{token} → deserialize JSON
Modify the field in application memory
SET session:{token} {updated JSON} → re-serialize and overwrite

Under concurrent requests (two requests updating different session fields simultaneously), one overwrites the other's changes. Race condition.

Option B: Redis Hash

text

HSET session:{token} userId 1001 email "j@example.com" role "engineer"
HSET session:{token} lastSeen "1717000000"
EXPIRE session:{token} 3600

Individual field updates are atomic:

text

HSET session:{token} lastSeen "1717003600"   → atomic, no read required
HINCRBY session:{token} pageViews 1           → atomic counter

Multiple fields can be updated without touching others. No race condition for independent field updates.

The trade-off: Hashes require one more key + the field name overhead. For sessions with 5–10 fields, the overhead is minimal. For sessions storing complex nested objects, you may need to serialize sub-objects as JSON strings within Hash fields.

Recommendation: Use a Hash for session storage. It enables atomic per-field updates and makes session data self-describing (HGETALL returns field names, not just a JSON blob).

typescript

// Session creation
async function createSession(userId: string, metadata: Record<string, string>) {
  const token = generateSecureToken();  // crypto.randomBytes(32).toString('hex')
  const key = `session:${token}`;

  await redis.hset(key,
    'userId', userId,
    'createdAt', String(Date.now()),
    'lastSeen', String(Date.now()),
    ...Object.entries(metadata).flat()
  );
  await redis.expire(key, 3600);  // 1-hour idle timeout

  return token;
}

// Session read (on every request)
async function getSession(token: string) {
  const data = await redis.hgetall(`session:${token}`);
  if (!data || Object.keys(data).length === 0) return null;
  return data;
}

Session Expiry: Idle Timeout vs Absolute Timeout

Idle Timeout (Sliding Expiry)

The session expires if unused for N minutes. Activity resets the timer.

typescript

async function touchSession(token: string, idleTimeoutSeconds = 3600): Promise<boolean> {
  const key = `session:${token}`;

  // Update lastSeen and reset TTL atomically
  const pipeline = redis.pipeline();
  pipeline.hset(key, 'lastSeen', String(Date.now()));
  pipeline.expire(key, idleTimeoutSeconds);
  const results = await pipeline.exec();

  // If the key didn't exist (expired between our check and touch), result is 0
  const expireResult = results?.[1]?.[1] as number;
  return expireResult === 1;
}

Call touchSession on every authenticated request. If it returns false, the session expired and the user should be logged out.

Absolute Timeout

The session expires at a fixed time after creation, regardless of activity. Implement by storing the absolute expiry timestamp in the session Hash and checking it on every access:

typescript

async function getSession(token: string) {
  const data = await redis.hgetall(`session:${token}`);
  if (!data || !data.userId) return null;

  const absoluteExpiry = parseInt(data.absoluteExpiry ?? '0', 10);
  if (absoluteExpiry && Date.now() > absoluteExpiry) {
    await redis.del(`session:${token}`);
    return null;
  }

  return data;
}

typescript

// On session creation: absolute expiry 24 hours from now
await redis.hset(key, 'absoluteExpiry', String(Date.now() + 86400 * 1000));

Combining Both

Most security-conscious systems use both:

Idle timeout: 30 minutes of inactivity
Absolute timeout: 24 hours regardless of activity

typescript

async function createSession(userId: string) {
  const token = generateSecureToken();
  const key = `session:${token}`;
  const now = Date.now();

  await redis.hset(key,
    'userId', userId,
    'createdAt', String(now),
    'lastSeen', String(now),
    'absoluteExpiry', String(now + 86400 * 1000)
  );
  await redis.expire(key, 1800);  // 30-minute idle timeout

  return token;
}

Multi-Device Session Management

Users log in from multiple devices. Each device gets its own session token. You need to:

List all active sessions for a user
Revoke a specific session (logout from one device)
Revoke all sessions (logout from all devices, e.g., after password change)

Track Sessions per User with a Set

typescript

async function createSession(userId: string, deviceInfo: string) {
  const token = generateSecureToken();
  const sessionKey = `session:${token}`;
  const userSessionsKey = `user:sessions:${userId}`;

  await redis.hset(sessionKey,
    'userId', userId,
    'device', deviceInfo,
    'createdAt', String(Date.now())
  );
  await redis.expire(sessionKey, 3600);

  // Track this session under the user's session index
  await redis.sadd(userSessionsKey, token);
  await redis.expire(userSessionsKey, 3600);  // clean up if user is inactive

  return token;
}

async function getActiveSessions(userId: string) {
  const tokens = await redis.smembers(`user:sessions:${userId}`);

  // Filter out expired sessions (their Hash keys won't exist)
  const sessions = await Promise.all(
    tokens.map(async (token) => {
      const data = await redis.hgetall(`session:${token}`);
      return Object.keys(data).length > 0 ? { token, ...data } : null;
    })
  );

  return sessions.filter(Boolean);
}

async function revokeSession(userId: string, token: string) {
  await redis.del(`session:${token}`);
  await redis.srem(`user:sessions:${userId}`, token);
}

async function revokeAllSessions(userId: string) {
  const tokens = await redis.smembers(`user:sessions:${userId}`);
  if (tokens.length > 0) {
    await redis.del(...tokens.map(t => `session:${t}`));
  }
  await redis.del(`user:sessions:${userId}`);
}

Memory cleanup: The user sessions Set may contain tokens of expired sessions (the session Hash expired but the token remains in the Set). The getActiveSessions function handles this by checking if the Hash exists, but the Set grows over time. Add a periodic cleanup or use SSCAN + EXISTS to prune stale tokens.

The Read-from-Replica Consistency Trap

Many Redis setups route read commands to replicas to reduce primary load. Session reads are a common target.

The problem: Replication in Redis is asynchronous. When a session is created or updated on the primary, the change propagates to replicas with a small delay (typically < 1ms, but up to seconds under load or network issues).

If a request is authenticated (session created on primary), and the next request reads the session from a replica before replication completes, the session lookup returns nil — the user appears logged out.

text

Request 1: POST /login  → session created on PRIMARY (token: abc123)
Response:  Set-Cookie: session=abc123

Request 2: GET /profile → reads from REPLICA (replica lag: 50ms)
           HGETALL session:abc123 → {} (not replicated yet)
           → user appears unauthenticated → 401 Unauthorized

Fix: Always read sessions from the primary.

typescript

// ioredis: use primary for session operations
const primaryRedis = new Redis({ host: 'redis-primary', port: 6379 });
const replicaRedis = new Redis({ host: 'redis-replica', port: 6379 });

// Sessions: always primary
async function getSession(token: string) {
  return primaryRedis.hgetall(`session:${token}`);
}

// Non-critical reads: can use replica
async function getProductData(productId: string) {
  return replicaRedis.get(`product:${productId}`);
}

For most session workloads, the primary handles the load comfortably — sessions are a small fraction of total Redis traffic.

Session Token Security

This is not Redis-specific, but engineers implementing Redis sessions often get this wrong:

Use a cryptographically secure random token:

typescript

import { randomBytes } from 'crypto';

function generateSecureToken(): string {
  return randomBytes(32).toString('hex');  // 64 hex characters, 256 bits of entropy
}

Do not use UUIDs (v4 UUIDs are random but some implementations have flaws; 122 bits of entropy vs 256).
Do not use incrementing integers or predictable patterns.
Do not sign session tokens with HMAC and put user data in the token unless you understand session forgery risks.

Set the cookie correctly:

Set-Cookie: session={token}; HttpOnly; Secure; SameSite=Strict; Path=/; Max-Age=3600

HttpOnly — JavaScript cannot access the cookie (XSS protection)
Secure — only sent over HTTPS
SameSite=Strict — CSRF protection

Summary

Use a Hash for session data — atomic per-field updates, self-documenting, no race conditions
Idle timeout via EXPIRE on each request; absolute timeout via a stored expiry field
Multi-device sessions: track token Set per user (user:sessions:{userId}); revoke by DEL session Hash + SREM from Set
Always read sessions from the primary — replica lag causes silent authentication failures
Secure tokens: 256-bit cryptographically random, stored in HttpOnly; Secure; SameSite=Strict cookies

Next: P-10 — Connection Pooling and Client Configuration — TCP connection overhead, ioredis pool sizing, reconnection strategies, and health check patterns for production clients.

Knowledge Check

A team decides to store user sessions as JSON strings in Redis (SET session:123 '{"userId":1, "lastSeen":170000}'). They have an endpoint that updates the lastSeen timestamp and another concurrent endpoint that adds a new item to a cart array within the same session. What is the fundamental risk of this approach compared to using a Redis Hash?

An application enforces both an "idle timeout" (30 minutes of inactivity) and an "absolute timeout" (24 hours maximum session life). Which implementation strategy is correct and most efficient?

To improve read performance, an engineering team configures their Node.js application to read session data from Redis read-replicas, while continuing to write new sessions to the primary. Shortly after, users report occasionally being "logged out" immediately after successfully logging in. What is the cause?

Test your knowledge with more question sets

PreviousModule P-8: Keyspace Notifications and Event-Driven Architectures Next Module P-10: Connection Pooling and Client Configuration

Discussion

Join the discussion

Loading comments...