Mussel — Airbnb’s Key-Worth Retailer for Derived Knowledge | by Shouyan guo | The Airbnb Tech Weblog

How Airbnb constructed a persistent, excessive availability and low latency key-value storage engine for accessing derived knowledge from offline and streaming occasions.

By: Chandramouli Rangarajan, Shouyan Guo, Yuxi Jin

Inside Airbnb, many on-line providers want entry to derived knowledge, which is knowledge computed with massive scale knowledge processing engines like Spark or streaming occasions like Kafka and saved offline. These providers require a top quality derived knowledge storage system, with sturdy reliability, availability, scalability, and latency ensures for serving on-line visitors. For instance, the person profiler service shops and accesses real-time and historic person actions on Airbnb to ship a extra customized expertise.

On this put up, we’ll discuss how we leveraged quite a few open supply applied sciences, together with HRegion, Helix, Spark, Zookeeper,and Kafka to construct a scalable and low latency key-value retailer for a whole lot of Airbnb product and platform use circumstances.

Over the previous few years, Airbnb has developed and enhanced our help for serving derived knowledge, transferring from groups rolling out customized options to a multi-tenant storage platform known as Mussel. This evolution will be summarized into three phases:

Stage 1 (01/2015): Unified read-only key-value retailer (HFileService)

Earlier than 2015, there was no unified key-value retailer resolution inside Airbnb that met 4 key necessities:

1. Scale to petabytes of knowledge
2. Environment friendly bulk load (batch era and importing)
3. Low latency reads (<50ms p99)
4. Multi-tenant storage service that can be utilized by a number of prospects

Additionally, not one of the current options had been capable of meet these necessities. MySQL doesn’t help bulk loading, Hbase’s huge bulk loading (distcp) isn’t optimum and dependable, RocksDB had no built-in horizontal sharding, and we didn’t have sufficient C++ experience to construct a bulk load pipeline to help RocksDB file format.

So we constructed HFileService, which internally used HFile (the constructing block of Hadoop HBase, which is predicated on Google’s SSTable):

1. Servers had been sharded and replicated to deal with scalability and reliability points
2. The variety of shards was mounted (equal to the variety of Hadoop reducers within the bulk load jobs) and the mapping of servers to shards saved in Zookeeper. We configured the variety of servers mapped to a selected shard by manually altering the mapping in Zookeeper
3. A each day Hadoop job reworked offline knowledge to HFile format and uploaded it to S3. Every server downloaded the information of their very own partitions to native disk and eliminated the previous variations of knowledge
4. Totally different knowledge sources had been partitioned by main key. Shoppers decided the proper shard their requests ought to go to by calculating the hash of the first key and modulo with the full variety of shards. Then queried Zookeeper to get an inventory of servers that had these shards and despatched the request to one in all them

Stage 2 (10/2015): Retailer each real-time and derived knowledge (Nebula)

Whereas we constructed a multi-tenant key-value retailer that supported environment friendly bulk load and low latency learn, it had its drawbacks. For instance, it didn’t help level, low-latency writes, and any replace to the saved knowledge needed to undergo the each day bulk load job. As Airbnb grew, there was an elevated must have low latency entry to real-time knowledge.

Subsequently, Nebula was constructed to help each batch-update and real-time knowledge in a single system. It internally used DynamoDB to retailer real-time knowledge and S3/HFile to retailer batch-update knowledge. Nebula launched timestamp based mostly versioning as a model management mechanism. For learn requests, knowledge can be learn from each an inventory of dynamic tables and the static snapshot in HFileService, and the consequence merged based mostly on timestamp.

To attenuate on-line merge operations, Nebula additionally had scheduled spark jobs that ran each day and merged snapshots of DynamoDB knowledge with the static snapshot of HFileService. Zookeeper was used to coordinate write availability of dynamic tables, snapshots being marked prepared for learn, and dropping of stale tables.

Stage 3 (2018): Scalable and low latency key-value storage engine (Mussel)

In Stage 3, we constructed a system that supported each learn and write on real-time and batch-update knowledge with timestamp-based battle decision. Nonetheless, there have been alternatives for enchancment:

1. Scale-out problem: It was cumbersome to manually edit partition mappings inside Zookeeper with rising knowledge progress, or to horizontally scale the system for rising visitors by including further nodes
2. Enhance learn efficiency underneath spiky write visitors
3. Excessive upkeep overhead: We would have liked to keep up HFileService and DynamoDB on the similar time
4. Inefficient merging course of: The method of merging the delta replace from DynamoDB and HFileService each day grew to become very gradual as our whole knowledge dimension grew to become bigger. The each day replace knowledge in DynamoDB was simply 1–2% of the baseline knowledge in HFileService. Nonetheless, we re-published the total snapshot (102% of whole knowledge dimension) again to HFileService each day

To resolve the drawbacks, we got here up with a brand new key-value retailer system known as Mussel.

1. We launched Helix to handle the partition mapping throughout the cluster
2. We leveraged Kafka as a replication log to copy the write to all the replicas as a substitute of writing on to the Mussel retailer
3. We used HRegion as the one storage engine within the Mussel storage nodes
4. We constructed a Spark pipeline to load the information from the information warehouse into storage nodes immediately

Let’s go into extra particulars within the following paragraphs.

Handle partitions with Helix

In Mussel, with a purpose to make our cluster extra scalable, we elevated the variety of shards from 8 in HFileService to 1024. In Mussel, knowledge is partitioned into these shards by the hash of the first keys, so we launched Apache Helix to handle these many logical shards. Helix manages the mapping of logical shards to bodily storage nodes robotically. Every Mussel storage node might maintain a number of logical shards. Every logical shard is replicated throughout a number of Mussel storage nodes.

Leaderless Replication with Kafka

Since Mussel is a read-heavy retailer, we adopted a leaderless structure. Learn requests could possibly be served by any of the Mussel storage nodes which have the identical logical shard, which will increase learn scalability. Within the write path, we wanted to think about the next:

1. We wish to clean the write visitors to keep away from the affect on the learn path
2. Since we don’t have the chief node in every shard, we’d like a manner to verify every Mussel storage node applies the write requests in the identical order so the information is constant throughout completely different nodes

To resolve these issues, we launched Kafka as a write-ahead-log right here. For write requests, as a substitute of immediately writing to the Mussel storage node, it’ll first write to Kafka asynchronously. Now we have 1024 partitions for the Kafka matter, every partition belonging to 1 logical shard within the Mussel. Every Mussel storage node will ballot the occasions from Kafka and apply the change to its native retailer. Since there isn’t a leader-follower relationship between the shards, this configuration permits the proper write ordering inside a partition, making certain constant updates. The downside right here is that it could solely present eventual consistency. Nonetheless, given the derived knowledge use case, it’s a suitable tradeoff to compromise on consistency within the curiosity of making certain availability and partition tolerance.

Supporting each learn, write, and compaction in a single storage engine

With the intention to scale back the {hardware} value and operational load of managing DynamoDB, we determined to take away it and lengthen HFileService as the one storage engine to serve each real-time and offline knowledge. To higher help each learn and write operations, we used HRegion as a substitute of Hfile. HRegion is a totally purposeful key-value retailer with MemStore and BlockCache. Internally it makes use of a Log Structured Merged (LSM) Tree to retailer the information and helps each learn and write operations.

An HRegion desk accommodates column households, that are the logical and bodily grouping of columns. There are column qualifiers inside a column household, that are the columns. Column households comprise columns with time stamped variations. Columns solely exist when they’re inserted, which makes HRegion a sparse database. We mapped our shopper knowledge to HRegion as the next:

With this mapping, for learn queries, we’re capable of help:

1. Level question by wanting up the information with main key
2. Prefix/vary question by scanning knowledge on secondary key
3. Queries for the newest knowledge or knowledge inside a selected time vary, as each real-time and offline knowledge written to Mussel could have a timestamp

As a result of we now have over 4000 shopper tables in Mussel, every person desk is mapped to a column household in HRegion as a substitute of its personal desk to cut back scalability challenges on the metadata administration layer. Additionally, as HRegion is a column-based storage engine, every column household is saved in a separate file to allow them to be learn/written independently.

For write requests, it consumes the write request from Kafka and calls the HRegion put API to put in writing the information immediately. For every desk, it could additionally help customizing the max model and TTL (time-to-live).

After we serve write requests with HRegion, one other factor to think about is compaction. Compaction must be run with a purpose to clear up knowledge that’s deleted or has reached max model or max TTL. Additionally when the MemStore in HRegion reaches a sure dimension, it’s flushed to disk right into a StoreFile. Compaction will merge these information collectively with a purpose to scale back disk search and enhance learn efficiency. Nonetheless, alternatively, when compaction is operating, it causes greater cpu and reminiscence utilization and blocks writes to stop JVM (Java Digital Machine) heap exhaustion, which impacts the learn and write efficiency of the cluster.

Right here we use Helix to mark Mussel storage nodes for every logical shard into two varieties of assets: on-line nodes and batch nodes. For instance, if we now have 9 Mussel storage nodes for one logical shard, 6 of them are on-line nodes and three of them are batch nodes. The connection between on-line and batch are:

1. They each serve write requests
2. Solely on-line nodes serve learn requests and we price restrict the compaction on on-line nodes to have good learn efficiency
3. Helix schedules a each day rotation between on-line nodes and batch nodes. Within the instance above, it strikes 3 on-line nodes to batch and three batch nodes to on-line so these 3 new batch nodes can carry out full pace main compaction to scrub up previous knowledge

With this modification, now we’re capable of help each learn and write with a single storage engine.

Supporting bulk load from knowledge warehouse

We help two varieties of bulk load pipelines from knowledge warehouse to Mussel by way of Airflow jobs: merge sort and substitute sort. Merge sort means merging the information from the information warehouse and the information from earlier write with older timestamps in Mussel. Substitute means importing the information from the information warehouse and deleting all the information with earlier timestamps.

We make the most of Spark to rework knowledge from the information warehouse into HFile format and add to S3. Every Mussel storage node downloads the information and makes use of HRegion bulkLoadHFiles API to load these HFiles into the column household.

With this bulk load pipeline, we will simply load the delta knowledge into the cluster as a substitute of the total knowledge snapshot on daily basis. Earlier than the migration, the person profile service wanted to load about 4TB knowledge into the cluster each day. After, it solely must load about 40–80GB, drastically decreasing the associated fee and bettering the efficiency of the cluster.

In the previous couple of years, Airbnb has come a good distance in offering a high-quality derived knowledge retailer for our engineers. The newest key-value retailer Mussel is extensively used inside Airbnb and has turn into a foundational constructing block for any key-value based mostly utility with sturdy reliability, availability, scalability, and efficiency ensures. Since its introduction, there have been ~4000 tables created in Mussel, storing ~130TB knowledge in our manufacturing clusters with out replication. Mussel has been working reliably to serve massive quantities of learn, write, and bulk load requests: For instance, mussel-general, our largest cluster, has achieved >99.9% availability, common learn QPS > 800k and write QPS > 35k, with common P95 learn latency lower than 8ms.

Despite the fact that Mussel can serve our present use circumstances effectively, there are nonetheless many alternatives to enhance. For instance, we’re wanting ahead to offering the read-after-write consistency to our prospects. We additionally wish to allow auto-scale and repartition based mostly on the visitors within the cluster. We’re wanting ahead to sharing extra particulars about this quickly.

Mussel is a collaborative effort of Airbnb’s storage group together with: Calvin Zou, Dionitas Santos, Ruan Maia, Wonhee Cho, Xiaomou Wang, Yanhan Zhang.

Desirous about engaged on the Airbnb Storage group? Try this position: Staff Software Engineer, Distributed Storage