# What is ICE?

The ICE (In-Memory Compute Engine) powers interactive analytics in Tellius. It keeps frequently used data hot in a vectorized columnar format so queries, search, and visualizations run fast.

**Purpose:** Low-latency, in-memory analytics.\
**Storage type:** Hot cache (memory/disk), not permanent storage.\
**Data flow:** Source → ICE (optimize/compress/partition) → Cache → Interactive compute → (optional) Publish to FQE.

### What it does

* Loads data in parallel from databases, cloud stores, and files.
* Stores and compresses data in a vectorized columnar format with fast in-memory compression for low-latency access.
* Shards and partitions data to maximize parallelism and data locality across the cluster.
* Caches data in memory, on local disk, or in a hybrid mode for predictable performance.
* Serves interactive workloads such as search, visualizations, transformations (ETL), Insights, and AutoML.
* Publishes denormalized Business Views to the query engine (FQE) for sub-second queries.

### How caching works

* **Memory cache:** Keeps optimized, compressed columns in RAM for the fastest response.
* **Disk cache:** Stores the same optimized format on local disks to extend capacity with strong performance.
* **Hybrid:** Uses a configurable portion of RAM and spills the rest to disk for balanced speed and scale.
* **Eviction policy:** Least-recently used datasets are evicted automatically to make space for new workloads.
* **Copy to System (optional):** Keeps a Parquet copy in distributed storage to accelerate cache rebuilds without re-hitting source systems.

### When to use ICE

* Interactive exploration with dashboards and search.
* Fast iteration on transformations and feature engineering.
* Running Insights and AutoML on hot data.
* Preparing denormalized Business Views for FQE.