Cache Management

Alluxio sits between your compute and persistent storage as a caching layer. Understanding how to manage that cache — what goes in, what stays, and what gets removed — is the key to consistent, high-performance workloads.

What do you want to do?

Goal	Tool	Guide
Pre-warm a dataset before a job runs	job load	Loading Data
Prevent a critical model from being evicted	Cache Priority (pinning)	Cache Policies → Pinning
Automatically expire stale or temporary data	TTL	Cache Policies → TTL
Control how fresh cached data must be	Cache Filter (maxAge)	Cache Policies → Filter
Limit cache usage per team or directory	Quota	Cache Policies → Quota
Remove a dataset after a job or version update	job free	Cache Eviction → Manual

How the Cache Lifecycle Works

1. Data Enters the Cache

Two paths:

• Passive (read-through): On a cache miss, the worker fetches from UFS and caches automatically. Zero configuration — but the first read is slow.

• Active preloading: Use job load to push data into cache before any read happens. Eliminates cold-start latency for scheduled jobs and model serving.

Cache Loading →

2. Policies Control What Stays and How Long

Once data is cached, four mechanisms govern its behavior:

• Cache Priority (Pinning) — marks data as HIGH priority so LRU never evicts it before lower-priority data

• TTL — sets a hard expiry: data is evicted when its lifetime expires, regardless of access

• Cache Filter — controls admission: immutable, skip cache, or max-age revalidation

• Quota — caps how much cache space a directory tree can consume

Cache Policies →

3. Data Leaves the Cache

Three ways data exits:

• Automatic eviction: LRU (default), LFU, or FIFO removes data when workers fill up

• TTL expiry: background scan removes data whose lifetime has elapsed

• Manual removal: job free explicitly purges a path — use this when invalidating a model version or freeing space on demand

Cache Eviction →

Strategy Guide

LLM / ML Model Serving

Examples: fine-tuned models, base model weights, embedding indexes.

Goal: Zero cold-start, critical models never evicted.

1. Preload before serving: job load --path s3://models/llama3/ --submit --verify

2. Pin against LRU: alluxio priority add --path s3://models/llama3/ --priority high

3. On model update: job free --path s3://models/llama3-v1/ --submit → reload v2

Periodically Updated Data

Examples: daily ETL output, retraining datasets.

Goal: Fresh data with good hit rate.

• Use maxAge filter (e.g. 1d) so Alluxio revalidates on next access after expiry

• Run job load --skip-if-exists after upstream update to pre-warm the new version

Temporary or Checkpoint Data

Examples: training checkpoints, temp query results.

Goal: Avoid filling cache with short-lived data.

• skipCache filter to bypass caching entirely, or

• Short TTL (e.g. 1h) + LOW priority so eviction hits this first

Compliance / Sensitive Data

Examples: PII logs, GDPR-scoped directories.

Goal: Hard limit on data lifetime in cache.

• TTL with a compliance-aligned window (e.g. 90d)

• job free immediately after processing completes