Redis prefetch cache with redis-rs

Implement a Redis prefetch cache in Rust with redis-rs

This guide shows you how to implement a Redis prefetch cache in Rust with the redis crate (redis-rs). It includes a small local web server built with axum and tokio so you can watch the cache pre-load at startup, see a background sync worker apply primary mutations within milliseconds, and break the cache to confirm that reads never fall back to the primary.

Overview

Prefetch caching pre-loads a working set of reference data into Redis before the first request arrives, so every read on the request path is a cache hit. A separate sync worker keeps the cache current as the source of truth changes — there is no fall-back to the primary on the read path.

That gives you:

  • Near-100% cache hit ratios for reference and master data
  • Sub-millisecond reads for lookup-heavy paths at peak traffic
  • All reference-data reads offloaded from the primary database
  • Source-database changes propagated into Redis within a few milliseconds
  • A long safety-net TTL that bounds memory if the sync pipeline ever stops

In this example, each cached category is stored as a Redis hash under a key like cache:category:{id}. The hash holds the category fields (id, name, display_order, featured, parent_id) and the key has a long safety-net TTL that the sync worker refreshes on every add or update event. Delete events remove the cache key outright, so there is no TTL to refresh in that case.

How it works

The flow has three independent paths:

  1. On startup, the demo server calls cache.bulk_load(primary.list_records().await).await, which pipelines DEL + HSET + EXPIRE for every record in one round trip.
  2. On every read, the application calls cache.get(entity_id).await, which runs HGETALL against Redis only. A miss is treated as an error, not a trigger to query the primary.
  3. On every primary mutation, the primary appends a change event to an in-process tokio::sync::mpsc channel. The sync worker task drains the channel and calls cache.apply_change(event).await. For an upsert, the helper rewrites the cache hash and refreshes the safety-net TTL; for a delete, it removes the cache key.

In a real system the in-process change channel is replaced by a CDC pipeline — Redis Data Integration, Debezium plus a lightweight consumer, or an equivalent tool that tails the source's binlog/WAL and pushes events into Redis.

The prefetch-cache helper

The PrefetchCache struct wraps the cache operations (source):

use redis::aio::ConnectionManager;
use redis::Client;

let client = Client::open("redis://localhost:6379/")?;
let conn = ConnectionManager::new(client).await?;

let cache = PrefetchCache::new(conn, CacheConfig::default());
let primary = MockPrimaryStore::new(80);

// Pre-load every primary record into Redis in one pipelined round trip.
let records = primary.list_records().await;
cache.bulk_load(records).await?;

// Start the sync worker so primary mutations propagate into Redis.
let sync = Arc::new(SyncWorker::new(primary.clone(), cache.clone()));
sync.start().await;

// Read paths now go to Redis only.
let result = cache.get("cat-001").await?;

The helper uses ConnectionManager, which is cheap to clone, auto-reconnects on failure, and is the recommended way to share a redis-rs connection across tokio tasks.

Data model

Each cached category is stored in a Redis hash:

cache:category:cat-001
  id            = cat-001
  name          = Beverages
  display_order = 1
  featured      = true
  parent_id     =

The implementation uses:

  • HSET + EXPIRE, pipelined, for the bulk load and every sync event
  • HGETALL on the read path
  • DEL for sync-delete events and explicit invalidation
  • SCAN to enumerate the cached keyspace and to clear the prefix
  • TTL to surface remaining safety-net time in the demo UI

Bulk load on startup

The bulk_load method pipelines a DEL + HSET + EXPIRE triple for every record. The pipeline is sent in a single round trip, so loading thousands of records takes one network RTT plus the time Redis spends executing the commands locally — typically tens of milliseconds even for a large reference table:

pub async fn bulk_load(
    &self,
    records: Vec<HashMap<String, String>>,
) -> RedisResult<u64> {
    let mut pipe = redis::pipe();
    let mut loaded: u64 = 0;
    for record in &records {
        let entity_id = match record.get("id") {
            Some(v) if !v.is_empty() => v.clone(),
            _ => continue,
        };
        let cache_key = self.cache_key(&entity_id);
        let pairs: Vec<(String, String)> =
            record.iter().map(|(k, v)| (k.clone(), v.clone())).collect();
        pipe.del(&cache_key).ignore();
        pipe.hset_multiple(&cache_key, &pairs).ignore();
        pipe.expire(&cache_key, self.cfg.ttl_seconds).ignore();
        loaded += 1;
    }
    if loaded > 0 {
        let mut conn = self.conn.clone();
        pipe.query_async::<_, ()>(&mut conn).await?;
    }
    self.stats.prefetched.fetch_add(loaded, Ordering::Relaxed);
    Ok(loaded)
}

A plain redis::pipe() (without .atomic()) is intentional on the startup path: nothing is reading the cache yet, the records do not need to be applied atomically as a set, and skipping MULTI/EXEC keeps the bulk load fast. The same method is used for the live /reprefetch reload, which is safe because the demo pauses the sync worker around the clear-and-reload sequence — see Re-prefetch under load below. If you call bulk_load directly from your own code on a cache that is already serving reads, either pause your writers first or switch to redis::pipe().atomic() so callers cannot observe a half-loaded record.

Reads from Redis only

The get method runs HGETALL and returns the cached hash. It does not fall back to the primary on a miss. In a healthy system, a miss never happens; if it does, the application surfaces it as an error and treats it as a sync-pipeline incident:

pub async fn get(&self, entity_id: &str) -> RedisResult<GetResult> {
    let cache_key = self.cache_key(entity_id);
    let mut conn = self.conn.clone();

    let started = Instant::now();
    let cached: HashMap<String, String> = conn.hgetall(&cache_key).await?;
    let redis_latency_ms = started.elapsed().as_secs_f64() * 1000.0;

    if !cached.is_empty() {
        self.stats.hits.fetch_add(1, Ordering::Relaxed);
        Ok(GetResult { record: Some(cached), hit: true, redis_latency_ms })
    } else {
        self.stats.misses.fetch_add(1, Ordering::Relaxed);
        Ok(GetResult { record: None, hit: false, redis_latency_ms })
    }
}

This is the key behavioural difference from cache-aside: the request path never touches the primary, so reference-data reads cannot contribute to primary database load.

Applying sync events

The sync worker calls apply_change for every primary mutation. For an upsert, the helper rewrites the cache hash and refreshes the safety-net TTL in one MULTI/EXEC pipeline so the cache never holds a stale mix of old and new fields. For a delete, it removes the cache key:

pub async fn apply_change(&self, change: &ChangeEvent) -> RedisResult<()> {
    let cache_key = self.cache_key(&change.id);
    let mut conn = self.conn.clone();

    match change.op.as_str() {
        "upsert" => {
            let fields = match &change.fields {
                Some(f) if !f.is_empty() => f,
                _ => return Ok(()),
            };
            let pairs: Vec<(String, String)> =
                fields.iter().map(|(k, v)| (k.clone(), v.clone())).collect();
            redis::pipe()
                .atomic()
                .del(&cache_key).ignore()
                .hset_multiple(&cache_key, &pairs).ignore()
                .expire(&cache_key, self.cfg.ttl_seconds).ignore()
                .query_async::<_, ()>(&mut conn).await?;
        }
        "delete" => {
            let _: i64 = conn.del(&cache_key).await?;
        }
        _ => {}
    }
    Ok(())
}

The DEL before the HSET ensures the cached hash contains exactly the fields the primary record has now — fields that have been dropped from the primary will not linger in Redis. The early return on empty fields matters: redis-rs panics if you call hset_multiple with an empty slice, so a malformed upsert with no fields is dropped (in a real CDC consumer you would route it to a dead-letter queue).

The sync worker

The SyncWorker runs a long-lived tokio::task that polls the primary's change channel with a short timeout. Every change is applied to Redis as soon as it arrives (source):

async fn run_loop(
    primary: Arc<MockPrimaryStore>,
    cache: PrefetchCache,
    poll_timeout: Duration,
    state: WorkerState,
) {
    loop {
        if *state.stop_rx.borrow() { return; }
        if *state.pause_rx.borrow() {
            state.idle_notify.notify_waiters();
            // park until the pause is lifted or the worker is stopped
            // ...
            continue;
        }
        match primary.next_change(poll_timeout).await {
            Some(change) => {
                if let Err(err) = cache.apply_change(&change).await {
                    eprintln!("[sync] failed to apply {:?} id={}: {}", change.op, change.id, err);
                }
            }
            None => continue,
        }
    }
}

In production this loop is replaced by a CDC consumer reading from RDI's Redis output stream, Debezium's Kafka topic, or an equivalent change feed. The shape stays the same: drain events, apply them to Redis, advance the consumer offset.

Invalidation and re-prefetch

Two helpers exist for testing and recovery:

  • invalidate(entity_id) deletes a single cache key. The demo uses it to simulate a sync-pipeline failure on one record.
  • clear() runs SCAN MATCH cache:category:* and deletes every key under the prefix. The demo uses it to simulate a full cache loss.

In both cases, the recovery path is to call bulk_load(primary.list_records().await).await again — re-prefetching from the primary. The demo exposes this as the "Re-prefetch" button so you can see the cache come back to a fully-warm state in one operation.

Re-prefetch under load

clear() and bulk_load() are not atomic against the sync worker. If a change event arrives between the snapshot (primary.list_records()) and the bulk write, the bulk write can overwrite a newer value; if a change event arrives between clear()'s SCAN and DEL, the cleared entry can immediately be recreated. The demo's /clear and /reprefetch handlers solve this by pausing the sync worker around the operation:

let _ = state.sync.pause(Duration::from_secs(2)).await;
let _ = state.cache.clear().await.unwrap_or(0);
let records = state.primary.list_records().await;
let loaded = state.cache.bulk_load(records).await.unwrap_or(0);
state.sync.resume().await;

pause() flips a tokio::sync::watch flag and blocks on a tokio::sync::Notify until the worker confirms it has parked. Change events that arrive during the pause sit in the primary's channel and apply in order once resume() is called, so no event is lost.

Hit/miss accounting

The helper keeps thread-safe atomic counters for hits, misses, prefetched records, sync events applied, and the average lag between a primary change and its application to Redis. The demo UI surfaces these so you can confirm the cache is absorbing all reads and the sync worker is keeping up:

pub fn stats(&self) -> serde_json::Value {
    let hits = self.stats.hits.load(Ordering::Relaxed);
    let misses = self.stats.misses.load(Ordering::Relaxed);
    // ...
    serde_json::json!({
        "hits": hits,
        "misses": misses,
        "hit_rate_pct": hit_rate_pct,
        "prefetched": prefetched,
        "sync_events_applied": sync_events_applied,
        "sync_lag_ms_avg": sync_lag_ms_avg,
    })
}

In production you would emit these as Prometheus counters and gauges. The sync-lag metric is the most important: a sudden rise indicates the CDC pipeline is falling behind.

Prerequisites

Before running the demo, make sure that:

  • Redis is running and accessible. By default, the demo connects to localhost:6379.
  • You have a working Rust toolchain (cargo 1.70+).
  • The demo's Cargo.toml pins the redis crate at 0.24+ with the tokio-comp, aio, and connection-manager features; cargo build will pull the rest of the dependency graph.

If your Redis server is running elsewhere, start the demo with --redis-host and --redis-port.

Running the demo

Get the source files

The demo consists of five files. Download them from the rust source folder on GitHub, or grab them with curl:

mkdir prefetch-cache-demo && cd prefetch-cache-demo
BASE=https://raw.githubusercontent.com/redis/docs/main/content/develop/use-cases/prefetch-cache/rust
curl -O $BASE/Cargo.toml
curl -O $BASE/cache.rs
curl -O $BASE/primary.rs
curl -O $BASE/sync_worker.rs
curl -O $BASE/demo_server.rs

Start the demo server

From that directory:

cargo run --bin demo_server

You should see something like:

Redis prefetch-cache demo server listening on http://127.0.0.1:8790
Using Redis at localhost:6379 with cache prefix 'cache:category:' and TTL 3600s
Prefetched 5 records in 90.1 ms; sync worker running

After starting the server, visit http://localhost:8790.

The demo accepts these flags:

  • --host / --port — HTTP bind host and port (defaults to 127.0.0.1:8790)
  • --redis-host / --redis-port — Redis location
  • --cache-prefix — cache key prefix (default cache:category:)
  • --ttl-seconds — safety-net TTL in seconds (default 3600)
  • --primary-latency-ms — simulated primary read latency (default 80)

It exposes a small interactive page where you can:

  • See which IDs are in the cache and in the primary, side by side
  • Read a category through the cache and confirm every read is a hit
  • Update a field on the primary and watch the sync worker rewrite the cache hash
  • Add and delete categories and watch them appear and disappear from the cache
  • Invalidate one key or clear the entire cache to simulate a sync-pipeline failure
  • Re-prefetch from the primary to recover from a broken cache state
  • Watch the average sync lag, and confirm primary reads stay at one until you re-prefetch — each /reprefetch adds another primary read for the snapshot, but normal request traffic never reaches the primary at all

The mock primary store

To make the demo self-contained, the example includes a MockPrimaryStore that stands in for a source-of-truth database (source):

pub struct MockPrimaryStore {
    pub read_latency_ms: u64,
    reads: AtomicU64,
    inner: Mutex<Inner>,                          // records + tx
    rx: Mutex<mpsc::UnboundedReceiver<ChangeEvent>>,
}

pub async fn update_field(&self, entity_id: &str, field: &str, value: &str) -> bool {
    let mut inner = self.inner.lock().await;
    // ... mutate the record ...
    emit_locked(&inner.tx, CHANGE_OP_UPSERT, entity_id, Some(snapshot));
    true
}

Every mutation appends a change event to an in-process tokio::sync::mpsc channel. The emit happens while the mutation lock is held, so two concurrent updates cannot interleave their channel sends — the queue order always matches the order in which the records map was mutated. The sync worker drains the channel with a 50 ms timeout via tokio::time::timeout and applies each event to Redis. In a real system this channel is replaced by a CDC pipeline — RDI on Redis Enterprise or Debezium with a Redis consumer on open-source Redis.

Production usage

This guide uses a deliberately small local demo so you can focus on the prefetch-cache pattern. In production, you will usually want to harden several aspects of it.

Replace the in-process change channel with a real CDC pipeline

The demo's in-process channel is the simplest possible stand-in for a CDC change feed. In production, the change feed lives outside the application process: an RDI pipeline configured against your primary database, Debezium connectors writing to Kafka or a Redis stream, or your application explicitly publishing change events from the write path. Whatever you choose, the consumer side stays the same — read events, apply them to Redis, advance the offset.

Use a long safety-net TTL, not a freshness TTL

The TTL on each cache key is a safety net: it bounds memory if the sync pipeline silently stops, so a stuck consumer cannot leave stale data in Redis indefinitely. The TTL is not the freshness mechanism — freshness comes from the sync worker, which refreshes the TTL on every add or update event (delete events remove the key). Pick a TTL that is comfortably longer than your worst-case sync lag plus your alerting window, so a transient sync hiccup never expires hot keys.

Decide what to do on a cache miss

A prefetch cache treats a miss as an error or a missing record. The two reasonable strategies are:

  • Return a 404 to the user. Appropriate when the cache is authoritative for the lookup — for example, when the user is asking for a category by ID and the ID is not in the cache.
  • Page on-call. A sustained miss rate on IDs you know exist is an incident: either the prefetch did not run, or the sync pipeline is broken.

Whichever you choose, do not fall back to the primary on the read path — that is what cache-aside is for, and conflating the two patterns breaks the load-isolation guarantee that prefetch provides.

Bound the working set to what fits in memory

Prefetch only works if the entire dataset fits in Redis memory with headroom. Estimate the size of your reference data, multiply by a growth factor, and confirm the result fits within your Redis instance's maxmemory minus what other use cases need. If the working set grows beyond what Redis can hold, switch the dataset to a cache-aside pattern instead — the request path will pay miss latency, but you will not OOM.

Reconcile periodically against the primary

CDC pipelines are eventually consistent: an event can be lost (broker outage, consumer crash, configuration drift) and the cache can silently diverge from the source. Run a periodic reconciliation job that re-reads all primary records, compares them against the cache, and either re-prefetches or fixes individual entries. Even running it once a day catches drift that ad-hoc inspection would miss.

Namespace cache keys in shared Redis deployments

If multiple applications share a Redis deployment, prefix cache keys with the application name (cache:billing:category:{id}) so different services cannot clobber each other's entries. CacheConfig::prefix takes the prefix exactly for this; the demo also accepts --cache-prefix on the command line.

Use ConnectionManager, not a bare connection

redis::aio::ConnectionManager is the recommended async connection type for redis-rs. It is cheap to clone (the demo passes a clone to every async task), reconnects automatically on transient failures, and pipelines commands behind the scenes. A bare redis::aio::Connection is not Clone and cannot be safely shared across tokio tasks. If you need a true connection pool — for example, to fan out heavy MULTI/EXEC traffic across multiple sockets — use deadpool-redis or bb8-redis.

Inspect cached entries directly in Redis

When testing or troubleshooting, inspect the stored cache keys directly to confirm the bulk load and the sync worker are writing what you expect:

redis-cli --scan --pattern 'cache:category:*'
redis-cli HGETALL cache:category:cat-001
redis-cli TTL cache:category:cat-001

If a key is missing for an ID that still exists in the primary, the prefetch did not run, the key expired without a sync refresh, or someone invalidated it. If a key is still present for an ID that was deleted in the primary, the delete event has not yet been applied. If the TTL is much lower than the configured safety-net value on a hot key, the sync worker is not keeping up.

Learn more

  • redis-rs crate - Install and use the Rust Redis client
  • HSET - Write hash fields
  • HGETALL - Read every field of a hash
  • EXPIRE - Set key expiration in seconds
  • DEL - Delete a key on invalidation or sync-delete
  • SCAN - Iterate the cached keyspace without blocking the server
  • TTL - Inspect remaining safety-net time on a key
  • Redis Data Integration - Configuration-driven CDC into Redis on Redis Enterprise and Redis Cloud
RATE THIS PAGE
Back to top ↑