여러 원본에 걸친 집계 파이프라인 실행하기

A cross-source aggregation pipeline in Atlas Data Federation runs the same aggregation stages you know from MongoDB, but the data under each stage may come from different physical systems — an S3 bucket, a live Atlas cluster, or both. The federated query engine handles all the routing, fan-out, and result merging transparently. From your application's perspective, it looks like a single MongoDB collection query.

The simplest cross-source query is a find() on a virtual collection backed by S3 files. The query engine reads the files, parses them, and applies the filter. Fields in the filter that match partition attributes in the path cause automatic file pruning. Fields that do not match partition attributes are applied as a post-read filter.

You can run any aggregation stage against S3-backed virtual collections: $match, $group, $project, $sort, $limit. The query engine pushes down stages where possible (especially $match for partition pruning and column pruning in Parquet) and executes remaining stages in its own compute layer after reading the data.

The most powerful cross-source pattern is using $lookup to join a live Atlas collection with an S3-archived collection. Start the pipeline from the live collection (the 'driver') and look up into the virtual S3-backed collection. Always place a $match early to minimize the number of lookups performed.

If your federated instance has multiple Atlas cluster stores, you can join collections from different Atlas clusters in a single pipeline. This is useful for multi-tenant or multi-region deployments where data is sharded across separate clusters and you need cross-cluster reports without merging clusters or building a separate reporting database.

After running a cross-source aggregation, you can write the results back to a live Atlas collection using $out or $merge. This is the ETL pattern: read historical data from S3, join with live data, compute aggregates, and write the results to a materialised view collection in Atlas that application queries can then read cheaply and quickly.

When querying Parquet files, Data Federation applies column pruning: if your $project stage specifies only certain fields, the query engine reads only those columns from the Parquet file (which stores data column-by-column). This can reduce bytes read by 90%+ compared to reading every column. Place your $project as early as possible in the pipeline for maximum column pruning benefit.

Atlas Data Federation logs query execution details in the Atlas UI under the Query History tab. Each query shows: bytes processed, execution time, and the number of files/partitions scanned. High bytes-processed numbers usually mean partition attributes are missing or the query does not match any partition keys. Use this log to tune your storage configuration and query patterns.

S3 files from different time periods or systems may have different schemas (field names, types, structure). Data Federation handles this gracefully — missing fields return null, extra fields are included. You can use $ifNull, $cond, and $convert in your pipeline to normalise varying schemas before grouping or joining.

Data Federation does not cache results between queries — each query re-reads the underlying sources. For dashboards that run the same report repeatedly, use the ETL pattern: schedule an aggregation that writes results to an Atlas collection via $merge, then have your dashboard query the fast Atlas collection. Atlas Triggers can schedule this refresh on any cron interval.

Be aware of current limitations: 1) Transactions are not supported on federated instances. 2) Index usage only applies to Atlas-backed collections, not S3 files. 3) Very large result sets may time out — use $out/$merge to write results instead of streaming them back. 4) Latency is higher than a live Atlas query due to S3 I/O — not suitable for user-facing, real-time queries.

In this lesson you learned: cross-source aggregation pipelines use the same MongoDB stages against virtual collections backed by S3 or Atlas clusters, $lookup enables joining live Atlas data with S3 archives in a single pipeline, and early $project enables column pruning in Parquet files to dramatically reduce bytes scanned. Next up we explore S3 data partitioning for query performance.