Constructing and deploying MySQL Raft at Meta


  • We’re rolling out MySQL Raft with the purpose to finally substitute our present MySQL semisynchronous databases. 
  • The most important win of MySQL Raft was simplification of the operation and making MySQL servers handle promotions and membership. This gave the provable security of Raft and diminished important operational ache.
  • Making MySQL server a real distributed system additionally has opened up potentialities in downstream methods to leverage it. A few of these concepts are beginning to take form.

At Meta, we run one of many largest deployments of MySQL on the earth. The deployment powers the social graph together with many different companies, like Messaging, Advertisements, and Feed. Over the previous few years, we’ve carried out MySQL Raft, a Raft consensus engine that was built-in with MySQL to construct a replicated state machine. We now have migrated a big portion of our deployment to MySQL Raft and plan to totally substitute the present MySQL semisynchronous databases with it. The undertaking has delivered important advantages to the MySQL deployment at Meta, together with larger reliability, provable security, important enhancements in failover time, and operational simplicity — all with equal or comparable write efficiency.


To permit for prime availability, fault tolerance, and scaling reads, Meta’s MySQL datastore is a massively sharded, geo-replicated deployment with thousands and thousands of shards, holding petabytes of knowledge. The deployment consists of hundreds of machines operating over a number of areas and information facilities throughout a number of continents.

Beforehand, our replication answer used the MySQL semisynchronous (semisync) replication protocol. This was an information path–solely protocol. The MySQL main would use semisynchronous replication to 2 log-only replicas (logtailers) inside the main area however exterior of the first’s failure area. These two logtailers would act as semisynchronous ACKer (An ACK is an acknowledgment to the first that the transaction has been regionally written).  This could permit the information path to have very low latency (sub-millisecond) commits and would offer excessive availability/sturdiness for the writes. Common MySQL primary-to-replica asynchronous replication was used for wider distribution to different areas.

The management airplane operations (e.g., promotions, failover, and membership change) can be the accountability of a set of Python daemons (henceforth referred to as automation). Automation would do the required orchestration to advertise a brand new MySQL server in a failover location as a main. The automation would additionally level the earlier main and the remaining replicas to copy from the brand new main. Membership change operations can be orchestrated by one other piece of automation referred to as MySQL pool scanner (MPS). So as to add a brand new member, MPS would level the brand new reproduction to the first and add it to the service discovery retailer. A failover can be a extra advanced operation during which the tailing threads of the logtailers (semisynchronous ACKers) can be shut right down to fence the earlier lifeless main.

Why was MySQL Raft crucial?

Previously, to assist assure security and keep away from information loss through the advanced promotion and failover operations, a number of automation daemons and scripts would use locking, orchestration steps, a fencing mechanism, and SMC,  a service discovery system. It was a distributed setup, and it was troublesome to perform this atomically. The automation turned extra advanced and tougher to keep up over time as increasingly more nook instances wanted to be patched.

We determined to take a very totally different method. We enhanced MySQL and made it a real distributed system. Realizing that management airplane operations like promotions and membership adjustments have been the set off of most points, we wished the management airplane and information airplane operations to be a part of the identical replicated log. For this, we used the well-understood consensus protocol Raft. This additionally meant that the supply of reality of membership and management moved contained in the server (mysqld). This was the only greatest contribution of bringing in Raft as a result of it enabled provable correctness (security property) throughout promotions and membership adjustments into the MySQL server.

The Raft library and the MySQL Raft plugin

Our implementation of Raft for MySQL is predicated on Apache Kudu. We enhanced it considerably for the wants of MySQL and our deployment. We revealed this fork as an open supply undertaking, kuduraft.

Among the key options that we added to kuduraft are:

  • FlexiRaft — assist for 2 totally different intersecting quorums: the information quorum and the chief election quorum
  • Proxying — the power to make use of a proxy intermediate node to scale back community bandwidth
  • Compression — the place we compress binary log (transaction) payloads as soon as earlier than distribution
  • Log abstraction — to assist totally different bodily logfile implementations
  • Major ban — the power to stop some entities from being main briefly

We additionally needed to make comparatively large adjustments to MySQL replication to interface with Raft. For this, we created a brand new closed supply MySQL plugin referred to as MyRaft. MySQL would interface with MyRaft by means of the plugin APIs (related APIs had been used for semisync as effectively), whereas we created a separate API for MyRaft to interface again with MySQL server (callbacks).

MySQL Raft replication topologies

A Raft ring would include a number of MySQL situations (4 in the diagram) in several areas. The communication round-trip time (RTT) between these areas would vary from 10 to 100 milliseconds. Just a few of those MySQLs (usually three) have been allowed to develop into primaries, whereas the remainder of them have been solely allowed to be pure learn replicas (non-primary-capable). The MySQL deployment at Meta additionally has a long-standing requirement for very low latency commits. The companies that use MySQL as a retailer (e.g., the social graph) want or have been designed to such extraordinarily quick writes.

To fulfill this requirement, the configuration of FlexiRaft would use solely in-region commits (single area dynamic mode). To allow this, every main succesful area would have two extra logtailers (witnesses or log-only entities). The information quorum for writes can be 2/3 (2 ACKs out of the 1 MySQL + 2 logtailers). Raft would nonetheless handle and run a replicated log throughout all of the entities (1 primary-capable MySQL + 2 logtailers ) * 3 areas + (non-primary-capable MySQL) * 3 areas = 12 entities.

Raft roles: The chief, because the title suggests, is the chief in a time period of the replicated log. A frontrunner in Raft would even be the first in MySQL and the one accepting shopper writes. The follower is a voting member of the ring and passively receives messages (AppendEntries) from the chief. A follower can be a reproduction in MySQL’s standpoint and can be making use of the transactions to its engine. It will not permit direct writes from person connections (read_only=1 is ready). A learner can be a non-voting member of the ring, e.g., the three MySQLs in non-primary-capable areas (above). It will be a reproduction in MySQL’s standpoint.

Replicated log

For replication, MySQL has traditionally used the binary log format. This format is central to MySQL’s replication, and we determined to protect this. From the Raft perspective, the binary log turned the replicated log. This was carried out through the log abstraction enchancment to kuduraft. The MySQL transactions can be encoded as a collection of occasions (e.g., Replace Rows occasion) with a begin and finish for every transaction. The binary log would even have applicable headers and would usually finish with an ending occasion (Rotate occasion). 

We needed to tweak how MySQL manages its logs internally. On a main, Raft would write to a binlog. That is no totally different from what occurs in commonplace MySQL. In a reproduction, Raft would additionally write to a binlog as an alternative of to a separate relay log in commonplace MySQL. This created simplicity for Raft as there was just one namespace of log information that Raft can be involved about. If a follower have been promoted to chief, it might seamlessly return into its historical past of logs to ship transactions to lagging members. The reproduction’s applier threads would choose up transactions from the binlog after which apply them to the engine. Throughout this course of, a brand new log file, the apply log, can be created. This apply log would play an vital function in crash restoration of replicas however is in any other case a nonreplicated log file.

So, in abstract:

In commonplace MySQL:

  • Major writes to binlog and sends binlog to replicas.
  • Replicas obtain in relay log and apply the transactions to the engine. Throughout apply, a brand new replica-only binlog is created.

In MySQL Raft:

  • Major writes to binlog through Raft, and Raft sends binlog to followers/replicas.
  • Replicas/followers obtain in binlog and apply the transactions to the engine. An apply log is created throughout apply.
  • Binlog is the replicated log from the Raft standpoint.

Write transaction on MySQL main utilizing Raft

The transaction would first be ready within the engine. This could occur within the thread of the person connection. The act of getting ready the transaction would contain interactions with the storage engine (e.g., InnoDB or MyRocks) and generate an in-memory binlog payload for the transaction. On the time of commit, the write would go by means of group commit/ordered_commit stream. GTIDs can be assigned, after which Raft would assign an OpId (time period:index) to the transaction. At this level, Raft would compress the transaction, retailer it in its LogCache, and write by means of the transaction to a binlog file. It will asynchronously begin delivery the transaction to different followers to get ACKs and attain consensus.

The person thread, which is in “commit” of the transaction, can be blocked, ready for consensus from Raft. When Raft would get two out of three in-region votes, consensus commit can be reached. Raft would additionally ship the transaction to all out-of-region members however would ignore their votes due to an algorithm referred to as FlexiRaft (described beneath). On consensus commit, the person thread can be unblocked, and the transaction would proceed and decide to the engine. After engine commit, the write question would end and return to the shopper. Quickly after, Raft would additionally asynchronously ship a commit marker (OpId of present commit) to downstream followers in order that they will additionally apply the transactions to their database. 

Crash restoration

Modifications needed to be made to crash restoration to make it work seamlessly with Raft. Crashes can occur at any time within the lifetime of a transaction and therefore the protocol has to make sure consistency of members. Listed here are some key insights on how we made it work.

  1. Transaction was not flushed to binlog: On this case, the in-memory transaction payload (nonetheless in mysqld course of reminiscence as an in-memory buffer) can be misplaced and the ready transaction in engine can be rolled again on course of restart. Since there was no additional uncommitted transaction within the Raft log, no reconciliation with different members must be carried out.
  2. Transaction was flushed to binlog however by no means reached different members: Mysqld acts as a transaction coordinator and runs a two-phase commit protocol between the engine and the replicated binlog because the members. On crash restoration, the ready transaction in engine (e.g., InnoDB or MyRocks) can be rolled again (engine had not reached commit). Raft would undergo failover, and a brand new chief can be elected. This chief wouldn’t have this transaction in its binlog and henceforth would truncate this transaction from the erstwhile chief’s binlog due to to a better time period (by pushing a No-Op message), when the erstwhile chief joins again the ring.
  3. Transaction was flushed to binlog and reached to subsequent chief. Present chief died earlier than committing to the engine: Just like no. 2 above, the ready transaction within the engine can be rolled again. The erstwhile chief would be part of the Raft ring as a follower. On this case, the brand new chief would have this transaction in its binlog and therefore no truncation would occur, because the logs would match. When the commit marker is distributed by the brand new chief, the transaction can be reapplied once more from scratch.

Raft-initiated state machine transitions

Failover and common upkeep operations can set off management adjustments in Raft. After a frontrunner is elected, the MyRaft plugin would attempt to to transition the accompanying MySQL into main mode. For this, the plugin would orchestrate a set of steps. These callbacks from Raft → MySQL would abort in-flight transactions, roll again in-use GTIDs, transition the engine aspect log from apply-log to binlog, and finally set the correct read_only settings. This mechanism is advanced and at present not open sourced.


Because the Raft paper and Apache Kudu supported solely a single world quorum, it will not work effectively at Meta, the place rings have been massive however the information path quorum wanted to be small.

To avoid this challenge, we innovated on FlexiRaft, borrowing concepts from Versatile Paxos.

At a excessive degree, FlexiRaft permits Raft to have a distinct information commit quorum (small) however take a corresponding hit on the chief election quorum (massive). By following provable ensures of quorum intersection, FlexiRaft ensures that the longest log guidelines of Raft and the suitable quorum intersection will assure provable security.

FlexiRaft helps single area dynamic mode. On this mode, members are grouped collectively by their geo-region. The present quorum of Raft will depend on who the present chief is (therefore the title “single area dynamic”). The information quorum is almost all of voters within the chief’s area. Throughout promotions, if phrases are steady, the Candidate will intersect with the final identified chief’s area. FlexiRaft would additionally be certain that the quorum of the Candidate’s area can also be attained, in any other case the following No-Op message might get caught. If within the uncommon case the phrases are usually not steady, Flexi Raft would strive to determine a rising set of areas which must be intersected with for security or, within the worst case, would fall again to the N area intersection case of Versatile Paxos. Because of pre-elections and mock elections, the incidences of time period gaps are uncommon.

Management airplane operations (promotions and membership adjustments)

In an effort to serialize promotion and membership change occasions within the binlog, we hijacked the Rotate Occasion and Metadata occasion of the MySQL binary log format. These occasions would carry the equal of No-Op messages and add-member/remove-member operations of Raft. Apache Kudu didn’t have assist for joint consensus, therefore we solely permit one-at-a-time membership adjustments (you may change the membership by just one entity in a single spherical to observe the foundations of implicit quorum intersection).


With the implementation of MySQL Raft, we reached a really clear separation of considerations for the MySQL deployment. The MySQL server can be accountable for security through Raft’s replicated state machine. The no-data-loss assure can be provably enshrined within the server itself. Automation (Python scripts, daemons) would provoke management airplane operations and monitor the well being of the fleet. It will additionally substitute members or do promotions through Raft throughout upkeep or when a number failure was detected. From time to time, automation might additionally change the regional placement of MySQL topology. Altering the automation to adapt to Raft was an enormous endeavor, spanning a number of years of improvement and rollout effort.

Throughout extended upkeep occasions, automation would set management ban info on Raft. Raft would disallow these banned entities from changing into chief or evacuate them promptly on inadvertent election. The automation would additionally promote away from these areas into different areas.

Studying from rollouts and challenges encountered alongside the best way

Rolling out Raft to the fleet was an enormous studying for the staff. We initially developed Raft on MySQL 5.6 and needed to migrate to MySQL 8.0.

One of many key learnings was that whereas correctness was simpler to cause with Raft, the Raft protocol in itself doesn’t assist a lot within the concern of availability. Since our MySQL information quorum was very small (two out of three in-region members), two dangerous entities within the area might just about shatter the quorum and produce down availability. The MySQL fleet undergoes quantity of churn every single day (because of upkeep, host failures, rebalancing operations), so initiating and doing membership adjustments promptly and appropriately have been a key necessities for fixed availability. A big a part of the rollout effort was centered at doing logtailer and MySQL replacements promptly in order that the Raft quorums have been wholesome.

We needed to improve kuduraft to make it extra sturdy for availability. These enhancements weren’t a part of the core protocol however might be thought-about as engineering add-ons to it. Kuduraft has the assist for pre-elections, however pre-elections are carried out solely throughout a failover. Throughout a sleek switch of management, the designated Candidate strikes on to an actual election, bumping the time period. This results in caught leaders (kuduraft doesn’t do auto step-down). To deal with this downside, we added a mock elections function, which was much like pre-elections however occurred solely upon a sleek switch of management. Since this was an async operation, it didn’t improve promotion downtimes. A mock election would weed out instances the place an actual election would partially succeed and get caught.

Dealing with byzantine failures: Raft’s membership checklist is taken into account to be blessed by Raft itself. However through the provisioning of latest members, or due to races in automation, there could possibly be weird instances of two totally different Raft rings intersecting. These zombie membership nodes needed to be weeded out and shouldn’t be capable of talk with one another. We carried out a function to dam RPCs from such zombie members to the ring. This was, in some methods, a dealing with of a byzantine actor. We enhanced the Raft implementation after noticing these uncommon incidents that occurred in our deployment.  

Monitoring the MySQL Raft rollout

Whereas launching MySQL Raft, one of many targets was to scale back operational complexity for on-calls, in order that engineers might root-cause and mitigate points. We constructed a number of dashboards, CLI instruments, and scuba tables to watch Raft. We added copious logging to MySQL, particularly across the space of promotions and membership adjustments. We created CLIs for quorum and voting stories on a hoop, which assist us rapidly determine when and why a hoop is unavailable (shattered quorum). The funding within the tooling and automation infrastructure went hand-in-hand and may need been a much bigger funding than the server adjustments. This funding paid off big-time and diminished operational and onboarding ache.

Quorum Fixer

Though it’s undesirable, quorums do get shattered from time to time, resulting in availability loss. The everyday case is when automation doesn’t detect unhealthy situations/logtailers within the ring and doesn’t substitute them rapidly. This will occur due to poor detection, employee queue overload, or a scarcity of spare host capability. Correlated failures, when a number of entities within the quorum go down on the identical time, are much less typical. These don’t occur typically, as a result of the deployments attempt to isolate failure domains throughout important entities of the quorum by means of correct placement selections. Lengthy story quick: At scale, surprising issues occur, regardless of current safeguards. Instruments must be obtainable to mitigate such conditions in manufacturing. We constructed Quorum Fixer in anticipation of this.

Quorum Fixer is a guide remediation device authored in Python that squelches the writes on the ring. It does out-of-band checks to determine the longest log entity. It forcibly adjustments the quorum expectations for a frontrunner election inside Raft, in order that the chosen entity turns into a frontrunner. After profitable promotion, we reset the quorum expectation again, and the ring usually turns into wholesome.

It was a aware determination to not run this device robotically, as a result of we wish to root trigger and determine all instances of quorum loss and repair bugs alongside the best way (not have them silently be mounted by automation).

Rolling out MySQL Raft

Transitioning from semisynchronous to MySQL Raft over an enormous deployment is troublesome. For this we created a device (in Python) referred to as enable-raft. Allow-raft orchestrates the transition from semisynchronous to Raft by loading the plugin and setting the suitable configs (mysql sys-vars) on every of the entities. This course of includes a small downtime for the ring. The device was made sturdy over time and may roll out Raft at scale in a short time. We now have used it to securely roll out Raft.

Testing and shadow workflow

For sure, doing a change within the core replication pipeline of MySQL is a really troublesome undertaking. Since information security is at stake, testing was key for confidence. We leveraged shadow testing and failure injection considerably through the undertaking. We’d inject hundreds of failovers and elections on check rings earlier than each RPM package deal supervisor rollout. We’d set off replacements and membership adjustments on the check property to set off the important code paths.

Lengthy-running testing with information correctness checks have been additionally key. We now have automation that runs nightly on the shards, making certain consistency of primaries and replicas. We’re alerted to any such mismatch, and we debug it.


The efficiency of the write path latency for Raft was equal to semisync. The semisync equipment is barely less complicated and therefore anticipated to be leaner, nevertheless we optimized Raft to get the identical latencies as semi-sync. We optimized kuduraft to not add any extra CPU to the fleet regardless of pulling in lots of extra duties that beforehand had been exterior the server binary.

Raft made order-of-magnitude enhancements to promotions and failover instances. Swish promotions, that are the majority of management adjustments within the fleet, improved considerably, and we are able to usually end a promotion in 300 milliseconds. Within the semisync setups, because the service discovery retailer can be the supply of reality, the purchasers noticing the end of promotion can be for much longer, resulting in extra elevated finish person downtimes on a shard.

Raft usually does a failover inside 2 seconds. It is because we heartbeat for Raft well being each 500 milliseconds and begin an election when three successive heartbeats fail. Within the semisync world, this step was orchestration heavy and would take 20 to 40 seconds. Raft thereby gave a 10x enchancment in downtimes for failover instances. 

Subsequent steps

Raft has helped resolve issues with the operational administration of MySQL at Meta by offering provable security and ease. Our targets of getting a arms off-management of MySQL consistency, and having instruments for the uncommon instances of availability loss, are largely met. Raft now opens up important alternatives sooner or later, as we are able to concentrate on enhancing the providing to the companies that use MySQL. One of many asks’ from our service house owners is to have configurable consistency. Configurable consistency will permit the house owners on the time of onboarding, to pick out whether or not the service wants X-region quorums or quorums that ask for copies in some particular geographies (e.g., Europe and the USA). FlexiRaft has seamless assist for such configurable quorums, and we plan to start out rolling out this assist sooner or later. Such quorums will correspondingly result in larger commit latencies, however use instances have to have the ability to trade-off between consistency and latency (e.g., PACELC theorem).

Due to the proxying function (potential to ship messages utilizing a multihop distribution topology), Raft may save community bandwidth throughout the Atlantic. We plan to make use of Raft to copy from the USA to Europe solely as soon as, after which use Raft’s proxying function to distribute inside Europe. It will improve latency, however it is going to be nominal on condition that the majority of the latency is within the cross-Atlantic switch and the additional hop is far shorter.

Among the extra speculative concepts in Meta’s database deployments and distributed consensus house are about exploring leaderless protocols, like Epaxos. Our present deployments and companies have labored with the assumptions that include sturdy chief protocols, however we’re beginning to see a trickle of necessities the place companies would profit from extra uniform write latency within the WAN. One other concept that we’re contemplating is to disentangle the log from the state machine (the database) right into a disaggregated log setup. It will permit the staff to handle the considerations of the log and replication individually from the considerations of the database storage and SQL execution engine.


Constructing and deploying MySQL Raft at Meta scale wanted important teamwork and administration assist. We wish to acknowledge the next individuals for his or her function in making this undertaking successful. Shrikanth Shankar, Tobias Asplund, Jim Carrig, Affan Dar and David Nagle for supporting the staff members throughout this journey. We’d additionally prefer to thank the ready Program Managers of this undertaking Dan O and Karthik Chidambaram who saved us on monitor.

The engineering effort concerned key contributions from a number of present and previous staff members together with Vinaykumar Bhat, Xi Wang, Bartlomiej Pelc, Chi Li, Yash Botadra, Alan Liang, Michael Percy, Yoshinori Matsunobu, Ritwik Yadav, Luqun Lou, Pushap Goyal, Anatoly Karp and Igor Pozgaj.