Garage is a data storage system that is based on CRDTs internally. It does not use a consensus algorithm such as Raft, therefore maintaining consistency in a cluster has to be done by other means. Since its inception, Garage has made use of read and write quorums to guarantee read-after-write consistency, the only consistency guarantee it provides. However, as of Garage v0.9.0, this guarantee is not maintained when the composition of a cluster is updated and data is moved between storage nodes. As part of our current NLnet-funded project, we are developing a solution to this problem. This blog post proposes a high-level overview of the proposed solution.
Garage provides mainly one consistency guarantee, read-after-write for objects, which can be described as follows:
Read-after-write consistency. If a client A writes an object x (e.g. using
PutObject) and receives a HTTP 200 OK response, and later a client B tries to
read object x (e.g. using GetObject), then B will read the version written by
A, or a more recent version.
The consistency guarantee offered by Garage is slightly more general than this simplistic formulation, as it also applies to other S3 endpoints such as ListObjects, which are always guaranteed to reflect the latest version of objects inserted in a bucket. Note that Amazon calls this guarantee strong read-after-write consistency (they also have it on AWS), to differentiate it from another definition of read-after-write consistency that only applies to data that is read by the same client that wrote it. Since that weaker form is also called read-your-writes, I will always be referring to the strong version when using the term "read-after-write consistency".
In Garage, this consistency guarantee at the level of objects in the S3 API is in fact a reflection of read-after-write consistency in the internal metadata engine (which is a distributed key/value store with CRDT values). Reads and writes to metadata tables use quorums of 2 out of 3 nodes for each operation, ensuring that if operation B starts after operation A has completed, then there is at least one node that is handling both operation A and B. In the case where A is a write (an update) and B is a read, that node will have the opportunity to return the value written in A to the reading client B. A visual depiction of this process can be found in this presentation on slide 32 (pages 57-64), and the algorithm is written down on slide 33 (page 54).
Note that read-after-write guarantees are broken and have always been for metadata related to buckets and access keys, which might not be something we can fix due to different requirements on the quorums for the related metadata tables.
Current issues with read-after-write consistency
Maintaining read-after-write consistency depends crucially on the intersection of the quorums being non-empty. There is however a scenario where these quorums may be empty: when the set of nodes affected to storing some entries changes, for instance when nodes are added or removed and data is being rebalanced between nodes.
A concrete example
Take the case of a partition (a subset of the data stored by Garage) which is stored on nodes A, B and C. At some point, a layout change occurs in the cluster, and after the change, nodes A, D and E are responsible for storing the partition. All read and write operations that were initiated before the layout change, or by nodes that were not yet aware of the new layout version, will be directed to nodes A, B and C, and will be handled by a quorum of two nodes among those three. However, once the new layout is introduced in the cluster, read and write operations will start being directed to nodes A, D and E, expecting a quorum of two nodes among this new set of three nodes.
Crucially, coordinating when operations start being directed to the new layout is a hard problem, and in all cases we must assume that due to some network asynchrony, there can still be some nodes that keep sending requests to nodes A, B and C for a long time even after everyone else is aware of the new layout. Moreover, data will be progressively moved from nodes B and C to nodes D and E, which can take a long time depending on the quantity of data. This creates a period of uncertainty as to where exactly the data is stored in the cluster. Overall, this basically means that this simplistic scheme gives us no way to guarantee the intersection-of-quorums property, which is necessary for read-after-write.
Concretely, here is a very simple scenario in which read-after-write is broken:
-
A write operation is directed to nodes A, B and C (the old layout), and receives OK responses from nodes B and C, forming a quorum, so the write completes successfully. The written data then arrives to node A as well.
-
The new layout version is introduced in the cluster.
-
Before nodes D and E have had the chance to retrieve the data that was stored on nodes B and C, a read operation for the same key is directed to nodes A, D and E. D and E both return an OK response with no data (a null value), because they is not yet up-to-date. An answer from node A is not received in time. The two responses from nodes D and E, that contain no data, still form a quorum, so the read returns a null value instead of the value that was written before, even though the write operation reported a success.
Evidencing the issue with Jepsen testing
The first thing that I had to do for the NLnet project was to develop a testing framework to show that read-after-write consistency issues could in fact arise in Garage when the cluster layout was updated. To make such tests, I chose to use the Jepsen testing framework, which helps us put distributed software in complex adverse scenarios and verify whether they respect some claimed consistency guarantees or not.
I will not enter into too much detail on the testing procedure, but suffice to say that issues were found. More precisely, I was able to show that Garage did guarantee read-after-write in a variety of adverse scenarios such as network partitions, node crashes and clock scrambling, but that it was unable to do so as soon as regular layout updates were introduced.
The progress of the Jepsen testing work is tracked in PR #544
Fixing read-after-write consistency when layouts change
To solve this issue, we will have to keep track of several pieces of information in the cluster. We will also have to adapt our read/write quorums and our data transfer strategy during rebalancing to make sure that data can be found when it is requested.
First of all, we adapted Garage's code to be able to handle several versions of the cluster layout that can be live in the cluster at the same time, to keep track of multiple possible locations for data that is currently being transferred between nodes. When multiple cluster layout versions are live, write operations are directed to all of the nodes responsible for storing the data in all the live versions. This ensures that the nodes in the oldest live layout version always have an up-to-date view of the data, and that a read quorum among those nodes is always a safe way to ensure read-after-write consistency.
Nodes will progressively synchronize data so that the nodes in the newest live layout version will catch up with data stored by nodes in the older live layout version. Once nodes in the newer layout versions also have an up-to-date view of the data, read operations will progressively start using a quorum of nodes in the new layout version instead of the old one.
Once all nodes are reading from newer layout versions, the oldest live versions can be pruned. This means that writes will stop being directed to those nodes, and the nodes will delete the data they were storing. Obviously, in the (very common) case where some nodes are both in the old and new layout versions, those nodes will not delete their data and they will continue to receive writes.
Performance impacts
When multiple layout versions are live, writes are sent to all nodes responsible for the partition of the requested key in all live layout versions, and will return OK only when they receive a quorum of OK responses for each of the live layout versions. This means that writes could be a bit slower when a layout change is being synchronized in the cluster. Typically if only one node is changing between the old and the new layout version, the write operation will await for 3 responses among 4 requests, instead of the classical 2 responses among 3 requests.
Concerning reads, they are still sent to only three nodes. Indeed, they are sent to the nodes of the newest live layout version for which nodes have completed a sync to catch up on existing data, and they only expect a quorum of 2 responses among the three nodes of that layout version. This way, reads always stay as performant as when no layout change is being processed.
Ensuring that new nodes are up-to-date
An additional coordination mechanism is necessary for the data synchronization procedure, to ensure that it is not started too early and that after it completes, the nodes in the new layout indeed contains an up-to-date view of the data.
Indeed, imagine the following adverse scenario, which we want to avoid: a new layout version is introduced in the cluster, and nodes immediately start copying the data to the new nodes. However, some write operations that were initiated before the new layout was introduced (or that were handled by a node not yet aware of the layout) could be delayed, and the written data was not yet received by the old nodes when they sent their copy of everything. When the sync reports completion, and read operations start being directed to nodes of the new layout, the written data might be missing from the nodes handling the read, and read-after-write consistency could be violated.
To avoid this situation, the synchronization operation is not initiated until all cluster nodes have reported an "acknowledge" of the new layout version, indicating that they have received the new layout version, and that they are no longer processing write operations that were only addressed to nodes of the previous layout versions. This makes sure that no data will be missed by the sync: once the sync has started, no more data can be written only to old layout versions. All of the writes will also be directed to the new nodes. More exactly: all data that the source nodes of the sync does not yet contain when the sync starts, is written by a write operation that is also directed at a quorum of nodes among the new ones. This means that at the end of the sync, a read quorum among the new nodes will necessarily return an up-to-date copy of all of the data.
Details on update trackers
As you can see, the previous algorithm needs to keep track of a lot of information in the cluster. This information is kept in three "layout update trackers", which keep track of the following information:
-
The
acklayout tracker keeps track of nodes receiving the latest layout versions and indicating that they are no longer processing writes addressed only to older layout versions. Once all nodes have acknowledged a new version, we know that all in-progress and future write operations that are made in the cluster are directed to the nodes that were added in this layout version as well. -
The
synclayout tracker keeps track of nodes finishing a full metadata table sync, that was started after all nodesack'ed the new layout version. -
The
sync_acklayout tracker keeps track of nodes receiving thesynctracker update for all cluster nodes, and thus starting to direct reads to the newly synchronized layout version. This makes it possible to know when no more nodes are reading from an old version, at which point the corresponding data can be deleted.
In the simplest scenario, only two layout versions are live, and these trackers
therefore can only have the values n (the new layout version) and n-1 (the
old one). However this mechanism handles the general case where several
successive layout updates are being processed and more than two layout versions
are live simultaneously. The layout update trackers can take as values the
version numbers of any currently live layout version.
What about dead nodes?
In this post I have used many times the phrases "once all nodes have acknowledged a new layout version", or "once all nodes have completed a sync". This obviously means that if some nodes are dead or unresponsive, the processing of the layout update can be delayed indefinitely, and nodes in the old layout versions will keep receiving writes and storing unnecessary data. This is an unfortunate fact with the method proposed here. To cover for these situations, the following workarounds can be made:
-
A layout change is generally a supervised operation, meaning that a system administrator may manually intervene to inform the cluster that certain nodes are dead and that their layout tracker values should not be taken into account.
-
For the
syncupdate tracker, we don't actually need to wait for all of the synchronizations to terminate, quorums can be used instead as they should be sufficient to ensure that the copied data is up-to-date. -
For the
ackandsync_ackupdate trackers, we can automatically increase them for all nodes (even dead ones) after a certain time delay, as there is no reason for the changes taking more than e.g. 10 minutes to propagate in regular conditions. We might not enable this behaviour by default, though, due to its possible impacts on consistency.
Current status and future work
The work described in this blog post is currently almost complete but it still needs to be ironed out. I have made a first run of Jepsen testing on the new code that showed that the changes seem to be fixing the issue. I will be running longer and more intensive runs of Jepsen testing once the code is finished, to make sure everything is fine. The changes will require a major update of Garage: this will be the v0.10.0 release, which will probably be finished in January or February of 2024. This update will be a very safe and transparent update, as only the layout data structure is changed and nothing related to object storage itself is touched.
If I had the time to do so, I would write the algorithm described in this post in a formal way, in the form of a scientific paper. I believe such a paper would be worthy of presenting at a scientific conference or journal, especially given the fact that it is motivated by a very concrete use case and has been validated quite thoroughly (with Jepsen). Unfortunately, this is not my highest priority at the moment.
Written by Alex Auvolat.