> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tell.rs/llms.txt
> Use this file to discover all available pages before exploring further.

# Sinks Overview

> Write data to ClickHouse, Vortex, Parquet, Arrow IPC, disk, or forward to another Tell instance.

Sinks receive data from the router and write it to storage. Each sink runs independently — if one is slow or down, the others continue unaffected.

## Available sinks

| Sink                                     | Type                            | Use case                                       |
| ---------------------------------------- | ------------------------------- | ---------------------------------------------- |
| [ClickHouse](/pipeline/sinks/clickhouse) | `clickhouse`                    | Production analytics (recommended)             |
| [Disk](/pipeline/sinks/disk)             | `disk_binary`, `disk_plaintext` | Binary or plaintext file storage               |
| [Parquet](/pipeline/sinks/parquet)       | `parquet`                       | Columnar archival with compression             |
| [Arrow IPC](/pipeline/sinks/arrow-ipc)   | `arrow_ipc`                     | Fast local storage for Polars/DuckDB           |
| [Vortex](/pipeline/sinks/vortex)         | `vortex`                        | Fast columnar reads with cascading compression |
| [Forwarder](/pipeline/sinks/forwarder)   | `forwarder`                     | Edge-to-cloud Tell-to-Tell relay               |
| Stdout                                   | `stdout`                        | Debug output (development only)                |
| Null                                     | `null`                          | Benchmarking (discards all data)               |

## Choosing a sink

**ClickHouse** is the recommended production sink. It handles event-type routing, per-table batching, concurrent flushes, and retry logic. The query engine connects directly to ClickHouse for analytics.

**Disk** sinks are for local file storage. Use `disk_binary` for high-throughput archival with optional LZ4 compression, or `disk_plaintext` for human-readable logs you can grep and tail.

**Parquet** is for data warehousing. Files are readable by Spark, DuckDB, Pandas, and Polars. Good for cold storage with excellent compression ratios.

**Arrow IPC** is for hot data that needs frequent access. \~10x faster reads than Parquet, with zero-copy memory mapping. Ideal for real-time dashboards backed by Polars or DuckDB.

**Vortex** is for fast analytical reads without a database. \~100x faster random access than Parquet via cascading columnar encodings. Use it when you need ClickHouse-class query speed on local files.

**Forwarder** sends data to another Tell instance over TCP. Use this for edge-to-cloud deployments where edge nodes collect data and relay it to a central server.

## Performance

Measured with `tell-bench sink` — 10M events, realistic cardinality (1000 devices, 25 event types, unique payloads), Apple M4 Pro:

| Sink                  | Events/s | Written | Ratio | Best for                        |
| --------------------- | -------- | ------- | ----- | ------------------------------- |
| disk\_binary          | 33.0M    | 639 MB  | 0.36x | Maximum write speed             |
| disk\_binary\_lz4     | 21.6M    | 145 MB  | 0.08x | Speed + compression             |
| arrow\_ipc            | 2.9M     | 1.7 GB  | 0.98x | Fast reads (Polars/DuckDB)      |
| parquet\_lz4          | 2.6M     | 251 MB  | 0.14x | Speed/compression balance       |
| parquet\_zstd         | 2.3M     | 171 MB  | 0.10x | Cold archival (best ratio)      |
| parquet\_uncompressed | 2.6M     | 769 MB  | 0.43x | Columnar without codec overhead |
| vortex                | 1.7M     | 1.8 GB  | 0.99x | Fast random access + scans      |
| disk\_plaintext       | 1.4M     | 2.7 GB  | 1.53x | Human-readable grep/tail        |

**Ratio** = bytes on disk / input bytes. Lower is better compression. Parquet's columnar layout eliminates per-event FlatBuffer overhead, so even "uncompressed" Parquet (0.43x) is smaller than raw binary (0.36x at full throughput).

## Backpressure

If a sink can't keep up, Tell drops batches for that sink rather than blocking the pipeline. Other sinks continue receiving data normally. Dropped batches are tracked in pipeline metrics — check `tell status --metrics` to spot sinks falling behind.

## File rotation

Disk-based sinks (disk, Parquet, Arrow IPC, Vortex) organize files by workspace and time:

```
{path}/{workspace_id}/{date}/{hour}/
```

Rotation can be `hourly` or `daily`. Disk sinks use atomic rotation to guarantee zero data loss during file switches.