Background & Motivation 

<WIP>

System Overview

<WIP>

Implementation

Writing 

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.

Index

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 dataset. 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:

Storage

The implementation specifics of the two storage types are detailed below.

Copy On Write (COW)

The Spark DAG for this storage, is relatively simpler. The key goal here is to group the tagged Hudi record RDD, into a series of updates and inserts, by using a partitioner. To achieve the goals of maintaining file sizes, we first sample the input to obtain a `workload profile` that understands the spread of inserts vs updates, their distribution among the partitions etc. With this information, we bin-pack the records such that

• For updates, the latest version of the that file id, is rewritten once, with new values for all records that have changed
• For inserts, the records are first packed onto the smallest file in each partition path, until it reaches the configured maximum size.

Any remaining records after that, are again packed into new file id groups, again meeting the size requirements. In this storage, index updation is a no-op, since the bloom filters are already written as a part of committing data. In the case of Copy-On-Write, a single parquet file constitutes one `file slice` which contains one complete version of the file

{% include image.html file="hudi_log_format_v2.png" alt="hudi_log_format_v2.png" max-width="1000" %}

Compactor

Realtime Readers will perform in-situ merge of these delta log-files to provide the most recent (committed) view of the dataset. To keep the query-performance in check and eventually achieve read-optimized performance, Hudi supports
compacting these log-files asynchronously to create read-optimized views.

Asynchronous Compaction involves 2 steps:

Compaction Schedule : Hudi Write Client exposes API to create Compaction plans which contains the list of `file slice` to be compacted atomically in a single compaction commit. Hudi allows pluggable strategies for choosing file slices for each compaction runs. This step is typically done inline by Writer process as Hudi expects only one schedule is being generated at a time which allows Hudi to enforce the constraint that pending compaction plans do not step on each other file-slices. This constraint allows for multiple concurrent `Compactors` to run at the same time. Some of the common strategies used for choosing `file slice` for compaction are:
BoundedIO - Limit the number of file slices chosen for a compaction plan by expected total IO (read + write) needed to complete compaction run
Log File Size - Prefer file-slices with larger amounts of delta log data to be merged
Day Based - Prefer file slice belonging to latest day partitions

Compactor : Hudi provides a separate API in Write Client to execute a compaction plan. The compaction plan (just like a commit) is identified by a timestamp. Most of the design and implementation complexities for Async Compaction is for guaranteeing snapshot isolation to readers and writer when multiple concurrent compactors are running. Typical compactor deployment involves launching a separate spark application which executes pending compactions when they become available. The core logic of compacting file slices in the Compactor is very similar to that of merging updates in a Copy-On-Write table. The only difference being in the case of compaction, there is an additional step of merging the records in delta log-files.

Here are the main API to lookup and execute a compaction plan.

Main API in HoodieWriteClient for running Compaction:
/**
* Performs Compaction corresponding to instant-time
* @param compactionInstantTime Compaction Instant Time
* @return
* @throws IOException
*/
public JavaRDD<WriteStatus> compact(String compactionInstantTime) throws IOException;

To lookup all pending compactions, use the API defined in HoodieReadClient

/**
* Return all pending compactions with instant time for clients to decide what to compact next.
* @return
*/
public List<Pair<String, HoodieCompactionPlan>> getPendingCompactions();
```
API for scheduling compaction

```

/**
* Schedules a new compaction instant
* @param extraMetadata
* @return Compaction Instant timestamp if a new compaction plan is scheduled
*/
Optional<String> scheduleCompaction(Optional<Map<String, String>> extraMetadata) throws IOException;

Refer to __hoodie-client/src/test/java/HoodieClientExample.java__ class for an example of how compaction is scheduled and executed.

Deployment Mode

These are typical Hudi Writer and Compaction deployment models

Inline Compaction : At each round, a single spark application ingests new batch to dataset. It then optionally decides to schedule a compaction run and executes it in sequence.
Single Dedicated Async Compactor: The Spark application which brings in new changes to dataset (writer) periodically schedules compaction. The Writer application does not run compaction inline. A separate spark applications periodically probes for pending compaction and executes the compaction.
Multi Async Compactors: This mode is similar to `Single Dedicated Async Compactor` mode. The main difference being now there can be more than one spark application picking different compactions and executing them in parallel.
In order to ensure compactors do not step on each other, they use coordination service like zookeeper to pickup unique pending compaction instants and run them.

The Compaction process requires one executor per file-slice in the compaction plan. So, the best resource allocation strategy (both in terms of speed and resource usage) for clusters supporting dynamic allocation is to lookup the compaction plan to be run to figure out the number of file slices being compacted and choose that many number of executors.

Async Compaction Design Deep-Dive 

For the purpose of this section, it is important to distinguish between 2 types of commits as pertaining to the file-group: A commit which generates a merged and read-optimized file-slice is called `snapshot commit` (SC) with respect to that file-group.  A commit which merely appended the new/updated records assigned to the file-group into a new log block is called `delta commit` (DC) with respect to that file-group.


The algorithm is described with an illustration. Let us assume a scenario where there are commits SC1, DC2, DC3 that have already completed on a data-set. Commit DC4 is currently ongoing with the writer (ingestion) process using it to upsert data. Let us also imagine there are a set of file-groups (FG1 … FGn) in the data-set whose latest version (`File-Slice`) contains the base file created by commit SC1 (snapshot-commit in columnar format) and a log file containing row-based log blocks of 2 delta-commits (DC2 and DC3).

{% include image.html file="async_compac_1.png" alt="async_compac_1.png" max-width="1000" %}

Writer (Ingestion) that is going to commit "DC4" starts. The record updates in this batch are grouped by file-groups and appended in row formats to the corresponding log file as delta commit. Let us imagine a subset of file-groups has
this new log block (delta commit) DC4 added. Before the writer job completes, it runs the compaction strategy to decide which file-group to compact by compactor and creates a new compaction-request commit SC5. This commit file is marked as “requested” with metadata denoting which fileIds to compact (based on selection policy). Writer completes without running compaction (will be run async).

{% include image.html file="async_compac_2.png" alt="async_compac_2.png" max-width="1000" %}

Writer job runs again ingesting next batch. It starts with commit DC6. It reads the earliest inflight compaction request marker commit in timeline order and collects the (fileId, Compaction Commit Id “CcId” ) pairs from meta-data. Ingestion DC6 ensures a new file-slice with base-commit “CcId” gets allocated for the file-group. The Writer will simply append records in row-format to the first log-file (as delta-commit) assuming the base-file (“Phantom-Base-File”) will be created eventually by the compactor.

{% include image.html file="async_compac_3.png" alt="async_compac_3.png" max-width="1000" %}

Compactor runs at some time and commits at “Tc” (concurrently or before/after Ingestion DC6). It reads the commit-timeline and finds the first unprocessed compaction request marker commit. Compactor reads the commit’s metadata finding the file-slices to be compacted. It compacts the file-slice and creates the missing base-file (“Phantom-Base-File”) with “CCId” as the commit-timestamp. Compactor then marks the compaction commit timestamp as completed. It is important to realize that at data-set level, there could be different file-groups requesting compaction at different commit timestamps.

{% include image.html file="async_compac_4.png" alt="async_compac_4.png" max-width="1000" %}

Near Real-time reader interested in getting the latest snapshot will have 2 cases. Let us assume that the incremental ingestion (writer at DC6) happened before the compaction (some time “Tc”’).  The below description is with regards to compaction from file-group perspective.  Reader querying at time between ingestion completion time for DC6 and compaction finish “Tc”`: Hudi’s implementation will be changed to become aware of file-groups currently waiting for compaction and merge log-files corresponding to DC2-DC6 with the base-file corresponding to SC1. In essence, Hudi will create a pseudo file-slice by combining the 2 file-slices starting at base-commits SC1 and SC5 to one.
For file-groups not waiting for compaction, the reader behavior is essentially the same - read latest file-slice and merge on the fly.

Reader querying at time after compaction finished (> “Tc”)` : In this case, reader will not find any pending compactions in the timeline and will simply have the current behavior of reading the latest file-slice and merging on-the-fly.

Read-Optimized View readers will query against the latest columnar base-file for each file-groups.

The above algorithm explains Async compaction w.r.t a single compaction run on a single file-group. It is important to note that multiple compaction plans can be run concurrently as they are essentially operating on different file-groups.