Repository analysis API

POST /_snapshot/<repository>/_analyze

There are a large number of third-party storage systems available, not all of which are suitable for use as a snapshot repository by Elasticsearch. Some storage systems behave incorrectly, or perform poorly, especially when accessed concurrently by multiple clients as the nodes of an Elasticsearch cluster do.

The Repository analysis API performs a collection of read and write operations on your repository which are designed to detect incorrect behaviour and to measure the performance characteristics of your storage system.

The default values for the parameters to this API are deliberately low to reduce the impact of running an analysis inadvertently and to provide a sensible starting point for your investigations. Run your first analysis with the default parameter values to check for simple problems. If successful, run a sequence of increasingly large analyses until you encounter a failure or you reach a blob_count of at least 2000, a max_blob_size of at least 2gb, a max_total_data_size of at least 1tb, and a register_operation_count of at least 100. Always specify a generous timeout, possibly 1h or longer, to allow time for each analysis to run to completion. Perform the analyses using a multi-node cluster of a similar size to your production cluster so that it can detect any problems that only arise when the repository is accessed by many nodes at once.

If the analysis fails then Elasticsearch detected that your repository behaved unexpectedly. This usually means you are using a third-party storage system with an incorrect or incompatible implementation of the API it claims to support. If so, this storage system is not suitable for use as a snapshot repository. You will need to work with the supplier of your storage system to address the incompatibilities that Elasticsearch detects. See Self-managed repository types for more information.

If the analysis is successful this API returns details of the testing process, optionally including how long each operation took. You can use this information to determine the performance of your storage system. If any operation fails or returns an incorrect result, this API returns an error. If the API returns an error then it may not have removed all the data it wrote to the repository. The error will indicate the location of any leftover data, and this path is also recorded in the Elasticsearch logs. You should verify yourself that this location has been cleaned up correctly. If there is still leftover data at the specified location then you should manually remove it.

If the connection from your client to Elasticsearch is closed while the client is waiting for the result of the analysis then the test is cancelled. Some clients are configured to close their connection if no response is received within a certain timeout. An analysis takes a long time to complete so you may need to relax any such client-side timeouts. On cancellation the analysis attempts to clean up the data it was writing, but it may not be able to remove it all. The path to the leftover data is recorded in the Elasticsearch logs. You should verify yourself that this location has been cleaned up correctly. If there is still leftover data at the specified location then you should manually remove it.

If the analysis is successful then it detected no incorrect behaviour, but this does not mean that correct behaviour is guaranteed. The analysis attempts to detect common bugs but it certainly does not offer 100% coverage. Additionally, it does not test the following:

The analysis comprises a number of blob-level tasks, as set by the blob_count parameter, and a number of compare-and-exchange operations on linearizable registers, as set by the register_operation_count parameter. These tasks are distributed over the data and master-eligible nodes in the cluster for execution.

For most blob-level tasks, the executing node first writes a blob to the repository, and then instructs some of the other nodes in the cluster to attempt to read the data it just wrote. The size of the blob is chosen randomly, according to the max_blob_size and max_total_data_size parameters. If any of these reads fails then the repository does not implement the necessary read-after-write semantics that Elasticsearch requires.

For some blob-level tasks, the executing node will instruct some of its peers to attempt to read the data before the writing process completes. These reads are permitted to fail, but must not return partial data. If any read returns partial data then the repository does not implement the necessary atomicity semantics that Elasticsearch requires.

For some blob-level tasks, the executing node will overwrite the blob while its peers are reading it. In this case the data read may come from either the original or the overwritten blob, but the read operation must not return partial data or a mix of data from the two blobs. If any of these reads returns partial data or a mix of the two blobs then the repository does not implement the necessary atomicity semantics that Elasticsearch requires for overwrites.

The executing node will use a variety of different methods to write the blob. For instance, where applicable, it will use both single-part and multi-part uploads. Similarly, the reading nodes will use a variety of different methods to read the data back again. For instance they may read the entire blob from start to end, or may read only a subset of the data.

For some blob-level tasks, the executing node will abort the write before it is complete. In this case it still instructs some of the other nodes in the cluster to attempt to read the blob, but all of these reads must fail to find the blob.

Linearizable registers are special blobs that Elasticsearch manipulates using an atomic compare-and-exchange operation. This operation ensures correct and strongly-consistent behavior even when the blob is accessed by multiple nodes at the same time. The detailed implementation of the compare-and-exchange operation on linearizable registers varies by repository type. Repository analysis verifies that that uncontended compare-and-exchange operations on a linearizable register blob always succeed. Repository analysis also verifies that contended operations either succeed or report the contention but do not return incorrect results. If an operation fails due to contention, Elasticsearch retries the operation until it succeeds. Most of the compare-and-exchange operations performed by repository analysis atomically increment a counter which is represented as an 8-byte blob. Some operations also verify the behavior on small blobs with sizes other than 8 bytes.