Redis Is Not Just a Cache

Redis is often described as "an in-memory key-value store used as a cache." This definition captures the most common use case but obscures it completely the real power of the system. Redis 7.x offers 10 types of native data structures, each optimized for specific access patterns – you don't have to model the data in reports or documents, you store them in the format that already reflects your queries.

What You Will Learn

  • Hash: Model structured objects with O(1) access to individual fields
  • Set and Sorted Set: sets, intersections and sorting by score
  • Real-time leaderboard with ZADD, ZRANK and ZRANGEBYSCORE
  • HyperLogLog: Approximate cardinality count with fixed 12KB
  • Geospatial: GEOADD, GEORADIUS and GEODIST for proximity search
  • When to use each structure and which to avoid

Hash: Structured Objects in Redis

A Redis Hash is a map of field-value pairs associated with a single key. It is the natural way to represent an object (a user, a product, a session) without serializing everything into JSON. The key advantage: you can read or update a single field in O(1) without loading the entire object. 

# Redis Hash: operazioni base
# HSET key field value [field value ...]
HSET user:1001 name "Mario Rossi" email "mario@example.com" age 35 city "Milano"

# HGET: singolo campo
HGET user:1001 name        # "Mario Rossi"
HGET user:1001 email       # "mario@example.com"

# HMGET: piu campi in una sola round trip
HMGET user:1001 name email city
# 1) "Mario Rossi"
# 2) "mario@example.com"
# 3) "Milano"

# HGETALL: tutto il hash (attenzione su hash grandi!)
HGETALL user:1001
# 1) "name"
# 2) "Mario Rossi"
# 3) "email"
# 4) "mario@example.com"
# 5) "age"
# 6) "35"
# 7) "city"
# 8) "Milano"

# HINCRBY: incremento atomico su campi numerici
HINCRBY user:1001 login_count 1

# HEXISTS: verifica esistenza campo
HEXISTS user:1001 phone     # 0 (non esiste)
HEXISTS user:1001 name      # 1

# HDEL: elimina campi specifici
HDEL user:1001 city

# HLEN: numero di campi
HLEN user:1001              # 4 (age, name, email, login_count)
# Python con redis-py
import redis

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

# Salva un oggetto utente
user_data = {
    'name': 'Mario Rossi',
    'email': 'mario@example.com',
    'age': '35',
    'city': 'Milano',
    'login_count': '0',
}
r.hset('user:1001', mapping=user_data)

# Lettura parziale: solo i campi che servono
name, email = r.hmget('user:1001', ['name', 'email'])
print(f"{name} <{email}>")  # Mario Rossi <mario@example.com>

# Incremento atomico del contatore login
new_count = r.hincrby('user:1001', 'login_count', 1)
print(f"Login count: {new_count}")  # 1

# Pattern: cache object con TTL
r.hset('session:abc123', mapping={
    'user_id': '1001',
    'role': 'admin',
    'created_at': '1710000000',
})
r.expire('session:abc123', 3600)  # 1 ora di TTL

# Lettura selettiva su campo singolo: O(1)
user_id = r.hget('session:abc123', 'user_id')  # '1001'

Sets: Unique Sets and Operations on Sets

A Redis Set is an unordered collection of unique strings. The strength of the Sets lies in the operations between sets: SUNION, SINTER, SDIFF calculate union, intersection and difference in server-side, without having to bring the data to the client. 

# Redis Set: tag system e operazioni su insiemi

# Aggiungi tag a articoli (SADD è idempotente per duplicati)
SADD article:100:tags "python" "fastapi" "backend"
SADD article:200:tags "python" "django" "web"
SADD article:300:tags "rust" "backend" "systems"

# SMEMBERS: tutti i membri (non ordinato)
SMEMBERS article:100:tags
# 1) "python"
# 2) "backend"
# 3) "fastapi"

# SISMEMBER: check membership O(1)
SISMEMBER article:100:tags "python"   # 1
SISMEMBER article:100:tags "java"     # 0

# SCARD: cardinalita del set
SCARD article:100:tags    # 3

# SINTER: articoli con tag in comune (python + backend)
SINTER article:100:tags article:300:tags
# 1) "backend"

# SUNION: tutti i tag di entrambi gli articoli
SUNION article:100:tags article:200:tags
# 1) "python"
# 2) "fastapi"
# 3) "backend"
# 4) "django"
# 5) "web"

# SDIFF: tag in article:100 ma NON in article:200
SDIFF article:100:tags article:200:tags
# 1) "fastapi"
# 2) "backend"

# SMOVE: sposta membro da un set all'altro (atomico)
SMOVE article:100:tags article:300:tags "backend"

# SRANDMEMBER: N elementi casuali (utile per sampling)
SRANDMEMBER article:100:tags 2

Sorted Set: Leaderboard and Range Queries

The Sorted Set is the most versatile data structure in Redis. Every element has one score associated float; the set is automatically sorted by score. You can read elements by position (rank) or by score range, all in O(log N). It is the natural choice for leaderboards, timelines, queues priority and range-based filtering. 

# Sorted Set: leaderboard gaming in tempo reale

# ZADD key score member
ZADD leaderboard:weekly 1500 "player:alice"
ZADD leaderboard:weekly 2300 "player:bob"
ZADD leaderboard:weekly 1800 "player:carol"
ZADD leaderboard:weekly 3100 "player:dave"
ZADD leaderboard:weekly 900  "player:eve"

# ZINCRBY: incremento atomico dello score (ogni kill += 100 punti)
ZINCRBY leaderboard:weekly 100 "player:alice"   # nuovo score: 1600

# ZRANK: posizione 0-based (dal basso, score crescente)
ZRANK leaderboard:weekly "player:bob"    # 2 (0-based)

# ZREVRANK: posizione dal top (score decrescente)
ZREVRANK leaderboard:weekly "player:dave"  # 0 (e' primo!)
ZREVRANK leaderboard:weekly "player:bob"   # 2

# ZSCORE: score di un membro specifico
ZSCORE leaderboard:weekly "player:carol"   # "1800"

# ZREVRANGE: top N giocatori (posizione, score decrescente)
ZREVRANGE leaderboard:weekly 0 4 WITHSCORES
# 1) "player:dave"
# 2) "3100"
# 3) "player:bob"
# 4) "2300"
# 5) "player:carol"
# 6) "1800"
# 7) "player:alice"
# 8) "1600"
# 9) "player:eve"
# 10) "900"

# ZRANGEBYSCORE: giocatori tra 1000 e 2000 punti
ZRANGEBYSCORE leaderboard:weekly 1000 2000 WITHSCORES
# 1) "player:alice"
# 2) "1600"
# 3) "player:carol"
# 4) "1800"

# ZCOUNT: quanti giocatori con score >= 1500
ZCOUNT leaderboard:weekly 1500 +inf    # 3

# ZCARD: totale membri nel sorted set
ZCARD leaderboard:weekly    # 5
# Python: leaderboard con Sorted Sets
import redis
from datetime import datetime

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

class Leaderboard:
    def __init__(self, name: str):
        self.key = f"leaderboard:{name}"

    def add_score(self, player: str, score: float) -> float:
        """Aggiunge punti al giocatore, restituisce nuovo totale."""
        return r.zincrby(self.key, score, player)

    def get_rank(self, player: str) -> int | None:
        """Posizione del giocatore (1-based, top = 1)."""
        rank = r.zrevrank(self.key, player)
        return rank + 1 if rank is not None else None

    def get_top(self, n: int = 10) -> list[dict]:
        """Top N giocatori con score."""
        entries = r.zrevrange(self.key, 0, n - 1, withscores=True)
        return [
            {'player': player, 'score': score, 'rank': i + 1}
            for i, (player, score) in enumerate(entries)
        ]

    def get_around(self, player: str, delta: int = 2) -> list[dict]:
        """I delta giocatori sopra e sotto un dato giocatore."""
        rank = r.zrevrank(self.key, player)
        if rank is None:
            return []
        start = max(0, rank - delta)
        end = rank + delta
        entries = r.zrevrange(self.key, start, end, withscores=True)
        return [
            {'player': p, 'score': s, 'rank': start + i + 1}
            for i, (p, s) in enumerate(entries)
        ]

# Uso
lb = Leaderboard("weekly")
lb.add_score("player:alice", 1500)
lb.add_score("player:bob", 2300)
lb.add_score("player:carol", 1800)
lb.add_score("player:dave", 3100)

print(lb.get_top(3))
# [{'player': 'player:dave', 'score': 3100.0, 'rank': 1}, ...]

print(lb.get_rank("player:carol"))  # 2
print(lb.get_around("player:bob", delta=1))  # bob + dave + carol

HyperLogLog: Cardinality Counting with Constant Memory

HyperLogLog is a probabilistic framework for estimating cardinality of a set (how many distinct elements there are) using an amount of memory constant: 12KB regardless of dataset size. The mistake standard is about 0.81%. He can't tell you which elements he saw, alone how many distinct. 

# HyperLogLog: conteggio unique views

# PFADD: aggiunge elementi all'HLL
PFADD page:article-100:views "user:alice" "user:bob" "user:carol"
PFADD page:article-100:views "user:alice"  # Duplicato: ignorato nella stima

# PFCOUNT: stima del numero di elementi distinti
PFCOUNT page:article-100:views    # 3 (stima, non esatto)

# Aggiunta in batch
PFADD page:article-100:views "user:dave" "user:eve" "user:frank"
PFCOUNT page:article-100:views    # 6

# PFMERGE: unisce piu HLL in uno (unique across multiple sets)
PFADD page:article-200:views "user:alice" "user:george" "user:henry"
PFMERGE all-articles page:article-100:views page:article-200:views
PFCOUNT all-articles    # ~8 (alice contata una sola volta nell'unione)

# Pattern: daily unique visitors
# Chiave per giorno: views:2026-03-20
PFADD views:2026-03-20 "user:alice"
PFADD views:2026-03-20 "user:bob"
# ... milioni di utenti, sempre 12KB

# Weekly count: merge dei 7 giorni
PFMERGE views:week-12 \
  views:2026-03-14 views:2026-03-15 views:2026-03-16 views:2026-03-17 \
  views:2026-03-18 views:2026-03-19 views:2026-03-20
PFCOUNT views:week-12    # Unique visitors della settimana
# Python: tracking unique page views con HyperLogLog
import redis
from datetime import date

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

def track_page_view(article_id: int, user_id: str) -> None:
    """Registra una view di un articolo da un utente."""
    today = date.today().isoformat()
    # Chiave giornaliera per articolo
    daily_key = f"hll:article:{article_id}:{today}"
    r.pfadd(daily_key, user_id)
    r.expire(daily_key, 90 * 86400)  # TTL 90 giorni

def get_unique_views(article_id: int, since_date: date, until_date: date) -> int:
    """Unique views in un range di date."""
    keys = []
    current = since_date
    while current <= until_date:
        keys.append(f"hll:article:{article_id}:{current.isoformat()}")
        current = date.fromordinal(current.toordinal() + 1)

    if not keys:
        return 0

    # Merge temporaneo per ottenere il conteggio dell'intero range
    temp_key = f"hll:temp:{article_id}:{since_date}:{until_date}"
    r.pfmerge(temp_key, *keys)
    r.expire(temp_key, 60)  # Cache il risultato per 60s
    return r.pfcount(temp_key)

# Track views
track_page_view(100, "user:alice")
track_page_view(100, "user:bob")
track_page_view(100, "user:alice")  # Secondo accesso: non conta

from datetime import date
views = get_unique_views(100, date(2026, 3, 1), date(2026, 3, 20))
print(f"Unique views marzo: ~{views}")

# Confronto memoria: Set vs HLL per 1 milione di utenti
# Redis Set: ~50MB (64 byte per elemento)
# HyperLogLog: 12KB fissi (4000x piu efficiente)

Geospatial: Proximity Searches with Geospatial Index

Redis Geospatial Index internally uses a Sorted Set where the score is a coordinate geohash. GEOADD, GEORADIUS and GEODIST allow you to do proximity search with complexity O(N + log M) where N is the number of results in the area and M the total of the elements. 

# Redis Geospatial: trova ristoranti vicini

# GEOADD key longitude latitude member
GEOADD restaurants 9.1859 45.4654 "ristorante-dal-mario"
GEOADD restaurants 9.1900 45.4680 "trattoria-lombarda"
GEOADD restaurants 9.1750 45.4600 "pizzeria-napoli"
GEOADD restaurants 9.2100 45.4800 "sushi-bento"
GEOADD restaurants 9.1850 45.4660 "bar-centrale"

# GEODIST: distanza tra due punti
GEODIST restaurants "ristorante-dal-mario" "trattoria-lombarda" km
# "0.3821" (circa 382 metri)

# GEOPOS: coordinate di un membro
GEOPOS restaurants "ristorante-dal-mario"
# 1) 1) "9.18589949607849121"
#    2) "45.46539883597492027"

# GEOSEARCH (Redis 6.2+): sostituisce GEORADIUS deprecato
# Trova ristoranti entro 500m dalla posizione corrente
GEOSEARCH restaurants
  FROMMEMBER "bar-centrale"
  BYRADIUS 500 m
  ASC
  COUNT 5
  WITHCOORD WITHDIST

# Oppure da coordinate GPS
GEOSEARCH restaurants
  FROMLONLAT 9.1860 45.4655
  BYRADIUS 1 km
  ASC
  COUNT 10

# GEOHASH: hash della posizione (per indicizzazione esterna)
GEOHASH restaurants "ristorante-dal-mario"
# 1) "u0nd0swfxh0"
# Python: proximity search per delivery app
import redis
from dataclasses import dataclass

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

@dataclass
class Restaurant:
    name: str
    longitude: float
    latitude: float
    category: str

def index_restaurant(restaurant: Restaurant) -> None:
    """Aggiunge ristorante all'indice geospaziale."""
    r.geoadd('restaurants:geo', {
        restaurant.name: (restaurant.longitude, restaurant.latitude)
    })
    # Salva metadata in un Hash separato
    r.hset(f"restaurant:{restaurant.name}", mapping={
        'name': restaurant.name,
        'category': restaurant.category,
        'lon': str(restaurant.longitude),
        'lat': str(restaurant.latitude),
    })

def find_nearby(lon: float, lat: float, radius_km: float, limit: int = 10) -> list[dict]:
    """Trova ristoranti entro radius_km dalla posizione."""
    results = r.geosearch(
        'restaurants:geo',
        longitude=lon,
        latitude=lat,
        radius=radius_km,
        unit='km',
        sort='ASC',
        count=limit,
        withdist=True,
        withcoord=True,
    )

    restaurants = []
    for entry in results:
        name, dist, (res_lon, res_lat) = entry
        metadata = r.hgetall(f"restaurant:{name}")
        restaurants.append({
            'name': name,
            'distance_km': round(dist, 3),
            'coordinates': {'lon': res_lon, 'lat': res_lat},
            'category': metadata.get('category', ''),
        })

    return restaurants

# Popola indice
for restaurant in [
    Restaurant("dal-mario", 9.1859, 45.4654, "italiana"),
    Restaurant("trattoria-lombarda", 9.1900, 45.4680, "italiana"),
    Restaurant("sushi-bento", 9.2100, 45.4800, "giapponese"),
]:
    index_restaurant(restaurant)

# Cerca ristoranti entro 1km da piazza Duomo Milano
nearby = find_nearby(lon=9.1895, lat=45.4654, radius_km=1.0)
for r_info in nearby:
    print(f"{r_info['name']}: {r_info['distance_km']}km ({r_info['category']})")

When to Use Which Data Structure

Quick Choice Guide

  • String: Simple values, counters, flags, JWT tokens, cached responses
  • Hash: Structured objects (users, sessions, products) with access to individual fields
  • List: FIFO/LIFO queues, activity feeds, logs sorted by insertion
  • Set: Tags, many-to-many relationships, membership checks, set operations
  • Sorted Set: Leaderboard, priority queue, ordered timeline, range queries
  • HyperLogLog: Approximate unique count (views, visitors) with constant memory
  • Geospatial: Proximity search, delivery radius, "near me" features
  • Bitmaps: Feature flags per user, daily presence tracking (DAU)
  • Streams: Persistent event log, message queue with consumer groups

Anti-Patterns to Avoid

# ANTI-PATTERN 1: HGETALL su hash enormi
# Se un hash ha 10.000 campi, HGETALL porta tutto in memoria del client
# Usa HSCAN per iterare in modo sicuro
cursor = 0
while True:
    cursor, fields = r.hscan('big-hash', cursor, count=100)
    # processa fields
    if cursor == 0:
        break

# ANTI-PATTERN 2: SMEMBERS su set molto grandi
# SMEMBERS blocca Redis per la durata della risposta
# Usa SSCAN invece
cursor = 0
while True:
    cursor, members = r.sscan('huge-set', cursor, count=100)
    # processa members
    if cursor == 0:
        break

# ANTI-PATTERN 3: KEYS * in produzione
# KEYS * blocca Redis finche non completa la scansione
# Usa SCAN con pattern
cursor = 0
while True:
    cursor, keys = r.scan(cursor, match='user:*', count=100)
    # processa keys
    if cursor == 0:
        break

# ANTI-PATTERN 4: Usare HLL quando hai bisogno dell'insieme esatto
# HLL non puo dirti QUALI elementi ha visto, solo QUANTI (approssimativamente)
# Se hai bisogno di sapere "questo utente ha visto questo articolo?",
# usa un Set o un Bloom Filter (RedisBloom)

Conclusions

The power of Redis is matching the right data structure to the problem. A Sorted Set for the leaderboard reduces application logic to zero: Redis maintains sorting automatically. A HyperLogLog for unique views use 12KB instead of 50MB. A Geospatial Index for proximity search eliminates the need for trigonometric calculations in the code. The next article will explore Pub/Sub and Streams, the two modes of Redis messaging with very different delivery guarantees.

Upcoming Articles in the Redis Series

  • Article 2: Pub/Sub vs Redis Streams — Consumer Groups and Distributed Processing
  • Article 3: Lua Scripting — Atomic Operations, EVAL and Redis Functions
  • Article 4: Rate Limiting — Token Bucket, Sliding Window and Fixed Counter