In this page hierarchy, we explain the concepts, design and the overall architectural underpinnings of Apache Hudi. This content is intended to be the technical documentation of the project and will be kept up-to date with 

Introduction

Apache Hudi (Hudi for short, here on) allows you to store vast amounts of data, on top existing def~hadoop-compatible-storage, while providing two primitives, that enable def~stream-processing on def~data-lakes, in addition to typical  def~batch-processing.

Specifically,

• Update/Delete Records : Hudi provides support for updating/deleting records, using fine grained file/record level indexes, while providing transactional guarantees for the write operation. Queries process  the last such committed snapshot, to produce results.
• Change Streams : Hudi also provides first-class support for obtaining an incremental stream of all the records that were updated/inserted/deleted in a given table, from a given point-in-time, and unlocks a new def~incremental-query category.

These primitives work closely hand-in-glove and unlock stream/incremental processing capabilities directly on top of def~DFS-abstractions. If you are familiar def~stream-processing, this is very similar to consuming events from a def~kafka-topic and then using a def~state-stores to accumulate intermediate results incrementally.

It has several architectural advantages.

• Increased Efficiency : Ingesting data often needs to deal with updates (resulting from def~database-change-capture), deletions (due to def~data-privacy-regulations) and enforcing def~unique-key-constraints (to ensure def~data-quality of event streams/analytics). However, due to lack of standardized support for such functionality using a system like Hudi, data engineers often resort to big batch jobs that reprocess entire day's events or reload the entire upstream database every run, leading to massive waste of def~computational-resources. Since Hudi supports record level updates, it brings an order of magnitude improvement to these operations, by only reprocessing changes records and rewriting only the part of the def~table, that was updated/deleted, as opposed to rewriting entire def~table-partitions or even the entire def~table.
• Faster ETL/Derived Pipelines : An ubiquitous next step, once the data has been ingested from external sources is to build derived data pipelines using Apache Spark/Apache Hive or any other data processing framework to  def~ETL the ingested data for a variety of use-cases like def~data-warehousing, def~machine-learning-feature-extraction, or even just def~analytics. Typically, such processes again rely on def~batch-processing jobs expressed in code or SQL, that process all input data in bulk and recompute all the output results. Such data pipelines can be sped up dramatically, by querying one or more input tables using an def~incremental-query instead of a regular def~snapshot-query, resulting once again in only processing the incremental changes from upstream tables and then def~upsert or delete the target derived table, like above.
• Access to fresh data :  It's not everyday, that you will find that reduced resource usage also result in improved performance, since typically we add more resources (e.g memory) to improve performance metric (e.g query latency) . By fundamentally shifting away from how datasets have been traditionally managed for may be the first time since the dawn of the big data era, Hudi, in fact, realizes this rare combination. A sweet side-effect of incrementalizing def~batch-processing is that the pipelines also much much smaller amount of time to run, putting data into hands of organizations much much quickly, than it has been possible with def~data-lakes before.
• Unified Storage : Building upon all the three benefits above, faster and lighter processing right on top of existing def~data-lakes mean lesser need for specialized storage or  def~data-marts, simply for purposes of gaining access to near real-time data.

System Overview

With an understanding of key technical motivations for the projects, let's now dive deeper into design of the system itself. At a high level,  components for writing Hudi tables are embedded into an Apache Spark job using one of the supported ways and it produces a set of files on def~backing-dfs-storage, that represents a Hudi def~table. Query engines like Apache Spark, Presto, Apache Hive can then query the table, with certain guarantees (that will discuss below).

There are three main components to a def~table

1. Set of  def~data-files that actually contain the records that were written to the table.
2. An def~index (which could be implemented in many ways), that maps a given record to a subset of the data-files that contains the record.
3. Ordered sequence of def~timeline-metadata about all the write operations done on the table, akin to a database transaction log.

Hudi provides the following capabilities for writers, queries and on the underlying data, which makes it a great building block for large def~data-lakes.

• upsert() support with fast, pluggable indexing
• Incremental queries that scan only new data efficiently
• Atomically publish data with rollback support, Savepoints for data recovery
• Snapshot isolation between writer & queries using def~mvcc style design
• Manages file sizes, layout using statistics
• Self managed def~compaction of updates/deltas against existing records.
• Timeline metadata to audit changes to data
• GDPR, Data deletions, Compliance.

Design Principles

Hudi is designed, from ground-up, for streaming records in and out of large datasets, borrowing principles from database design and def~log-structured-storage systems. To that end, Hudi provides def~index implementations, that can quickly map a record's key to the file location it resides at. Similarly, for streaming data out, Hudi adds and tracks record level metadata via def~hoodie-special-columns, that enables providing a precise incremental stream of all changes that happened. 

Hudi recognizes the different expectation of data freshness (write friendly) vs query performance (read/query friendliness) users may have, and supports three different def~query-types that provide real-time snapshots, incremental streams or purely columnar data that slightly older. At each step, Hudi strives to be self-managing (e.g: autotunes the writer parallelism, maintains file sizes) and self-healing (e.g: auto rollbacks failed commits), even if it comes at cost of slightly additional runtime cost (e.g: caching input data in memory to profile the workload). The core premise here, is that, often times operational costs of these large data pipelines without such operational levers/self-managing features built-in, dwarf the extra memory/runtime costs associated.

Hudi also has an append-only, cloud data storage friendly design, that lets Hudi manage data on across all the major cloud providers seamlessly.

Timeline

At its core, Hudi maintains a timeline of all def~instant-action performed on the def~table at different instants of time that helps provide instantaneous views of the def~table, while also efficiently supporting retrieval of data in the order in which it was written. The timeline is akin to a redo/transaction log, found in databases, and consists of a set of def~timeline-instants. Hudi guarantees that the actions performed on the timeline are atomic & timeline consistent based on the instant time. Timeline is implemented as a set of files under the `.hoodie` def~metadata-folder directly under the def~table-basepath. Specifically, while the most recent instants are maintained as individual files, the older instants are archived to the def~timeline-archival folder, to bound the number of files, listed by writers and queries. 

Storage/Writing

The implementation specifics of the two def~table-types are detailed below.

Copy On Write 

def~copy-on-write

Merge On Read

def~merge-on-read

Hudi writing is implemented as a Spark library, which makes it easy to integrate into existing data pipelines or ingestion libraries (which we will refer to as `Hudi clients`). Hudi Clients prepare an `RDD[HoodieRecord]` that contains the data to be upserted and Hudi upsert/insert is merely a Spark DAG, that can be broken into two big pieces.

• Indexing : A big part of Hudi's efficiency comes from indexing the mapping from record keys to the file ids, to which they belong to. This index also helps the `HoodieWriteClient` separate upserted records into inserts and updates, so they can be treated differently. `HoodieReadClient` supports operations such as `filterExists` (used for de-duplication of table) and an efficient batch `read(keys)` api, that can read out the records corresponding to the keys using the index much quickly, than a typical scan via a query. The index is also atomically updated each commit, and is also rolled back when commits are rolled back.
• Storage : The storage part of the DAG is responsible for taking an `RDD[HoodieRecord]`, that has been tagged as an insert or update via index lookup, and writing it out efficiently onto storage.

File Layout

<WIP>

Indexing

<WIP>

Hudi currently provides two choices for indexes : `BloomIndex` and `HBaseIndex` to map a record key into the file id to which it belongs to. This enables us to speed up upserts significantly, without scanning over every record in the table. Hudi Indices can be classified based on their ability to lookup records across partition. A `global` index does not need partition information for finding the file-id for a record key but a `non-global` does.

HBase Index (global)

Here, we just use HBase in a straightforward way to store the mapping above. The challenge with using HBase (or any external key-value store for that matter) is performing rollback of a commit and handling partial index updates.
Since the HBase table is indexed by record key and not commit Time, we would have to scan all the entries which will be prohibitively expensive. Instead, we store the commit time with the value and discard its value if it does not belong to a valid commit.

Bloom Index (non-global)

This index is built by adding bloom filters with a very high false positive tolerance (e.g: 1/10^9), to the parquet file footers. The advantage of this index over HBase is the obvious removal of a big external dependency, and also nicer handling of rollbacks & partial updates since the index is part of the data file itself.

At runtime, checking the Bloom Index for a given set of record keys effectively amounts to checking all the bloom filters within a given partition, against the incoming records, using a Spark join. Much of the engineering effort towards the Bloom index has gone into scaling this join by caching the incoming RDD[HoodieRecord] and dynamically tuning join parallelism, to avoid hitting Spark limitations like 2GB maximum for partition size. As a result, Bloom Index implementation has been able to handle single upserts upto 5TB, in a reliable manner.

DAG with Range Pruning:

Compaction

<WIP>

File Sizing

<WIP>

Querying

<WIP>

Snapshot Queries

<WIP>

Incremental Queries

<WIP>

Read Optimized

<WIP>