Snapshots of Unchanged Shards Could be Optimized #45736
Assignees
Labels
:Distributed Coordination/Snapshot/Restore
Anything directly related to the `_snapshot/*` APIs
>enhancement
Team:Distributed (Obsolete)
Meta label for distributed team (obsolete). Replaced by Distributed Indexing/Coordination.
Comments
original-brownbear
added
>enhancement
:Distributed Coordination/Snapshot/Restore
|
Pinging @elastic/es-distributed |
original-brownbear
added a commit
that referenced
this issue
Oct 22, 2019
### Changes to Root-Level index-N (RepositoryData)
This change adds a new field `"shards"` to `RepositoryData` that contains a mapping of `IndexId` to a `String[]`. This string array can be accessed by shard id to get the generation of a shard's shard folder (i.e. the `N` in the name of the currently valid `/indices/${indexId}/${shardId}/index-${N}` for the shard in question).
### Benefits
This allows for creating a new snapshot in the shard without doing any LIST operations on the shard's folder. In the case of AWS S3, this saves about 1/3 of the cost for updating an empty shard (see #45736) and removes one out of two remaining potential issues with eventually consistent blob stores (see #38941 ... now only the root `index-${N}` is determined by listing).
Also and equally if not more important, a number of possible failure modes on eventually consistent blob stores like AWS S3 are eliminated by moving all delete operations to the `master` node and moving from incremental naming of shard level index-N to uuid suffixes for these blobs.
### Only Master Deletes Blobs
This change moves the deleting of the previous shard level `index-${uuid}` blob to the master node instead of the data node allowing for a safe and consistent update of the shard's generation in the `RepositoryData` by first updating `RepositoryData` and then deleting the now unreferenced `index-${newUUID}` blob.
__No deletes are executed on the data nodes at all for any operation with this change.__
Note also: Previous issues with hanging data nodes interfering with master nodes are completely impossible, even on S3 (see next section for details).
### Why Move from index-${N} to index-${uuid} at the Shard Level
This change changes the naming of the shard level `index-${N}` blobs to a uuid suffix `index-${UUID}`. The reason for this is the fact that writing a new shard-level `index-` generation blob is not atomic anymore in its effect. Not only does the blob have to be written to have an effect, it must also be referenced by the root level `index-N` (`RepositoryData`) to become an effective part of the snapshot repository.
This leads to a problem if we were to use incrementing names like we did before. If a blob `index-${N+1}` is written but due to the node/network/cluster/... crashes the root level `RepositoryData` has not been updated then a future operation will determine the shard's generation to be `N` and try to write a new `index-${N+1}` to the already existing path. Updates like that are problematic on S3 for consistency reasons, but also create numerous issues when thinking about stuck data nodes.
Previously stuck data nodes that were tasked to write `index-${N+1}` but got stuck and tried to do so after some other node had already written `index-${N+1}` were prevented form doing so (except for on S3) by us not allowing overwrites for that blob and thus no corruption could occur.
Were we to continue using incrementing names, we could not do this. The stuck node scenario would either allow for overwriting the `N+1` generation or force us to continue using a `LIST` operation to figure out the next `N` (which would make this change pointless).
With uuid naming and moving all deletes to `master` this becomes a non-issue. Data nodes write updated shard generation `index-${uuid}` and `master` makes those `index-${uuid}` part of the `RepositoryData` that it deems correct and cleans up all those `index-` that are unused.
Co-authored-by: Yannick Welsch <[email protected]>
Co-authored-by: Tanguy Leroux <[email protected]>
original-brownbear
added a commit
to original-brownbear/elasticsearch
that referenced
this issue
Oct 23, 2019
This change adds a new field `"shards"` to `RepositoryData` that contains a mapping of `IndexId` to a `String[]`. This string array can be accessed by shard id to get the generation of a shard's shard folder (i.e. the `N` in the name of the currently valid `/indices/${indexId}/${shardId}/index-${N}` for the shard in question).
This allows for creating a new snapshot in the shard without doing any LIST operations on the shard's folder. In the case of AWS S3, this saves about 1/3 of the cost for updating an empty shard (see elastic#45736) and removes one out of two remaining potential issues with eventually consistent blob stores (see elastic#38941 ... now only the root `index-${N}` is determined by listing).
Also and equally if not more important, a number of possible failure modes on eventually consistent blob stores like AWS S3 are eliminated by moving all delete operations to the `master` node and moving from incremental naming of shard level index-N to uuid suffixes for these blobs.
This change moves the deleting of the previous shard level `index-${uuid}` blob to the master node instead of the data node allowing for a safe and consistent update of the shard's generation in the `RepositoryData` by first updating `RepositoryData` and then deleting the now unreferenced `index-${newUUID}` blob.
__No deletes are executed on the data nodes at all for any operation with this change.__
Note also: Previous issues with hanging data nodes interfering with master nodes are completely impossible, even on S3 (see next section for details).
This change changes the naming of the shard level `index-${N}` blobs to a uuid suffix `index-${UUID}`. The reason for this is the fact that writing a new shard-level `index-` generation blob is not atomic anymore in its effect. Not only does the blob have to be written to have an effect, it must also be referenced by the root level `index-N` (`RepositoryData`) to become an effective part of the snapshot repository.
This leads to a problem if we were to use incrementing names like we did before. If a blob `index-${N+1}` is written but due to the node/network/cluster/... crashes the root level `RepositoryData` has not been updated then a future operation will determine the shard's generation to be `N` and try to write a new `index-${N+1}` to the already existing path. Updates like that are problematic on S3 for consistency reasons, but also create numerous issues when thinking about stuck data nodes.
Previously stuck data nodes that were tasked to write `index-${N+1}` but got stuck and tried to do so after some other node had already written `index-${N+1}` were prevented form doing so (except for on S3) by us not allowing overwrites for that blob and thus no corruption could occur.
Were we to continue using incrementing names, we could not do this. The stuck node scenario would either allow for overwriting the `N+1` generation or force us to continue using a `LIST` operation to figure out the next `N` (which would make this change pointless).
With uuid naming and moving all deletes to `master` this becomes a non-issue. Data nodes write updated shard generation `index-${uuid}` and `master` makes those `index-${uuid}` part of the `RepositoryData` that it deems correct and cleans up all those `index-` that are unused.
Co-authored-by: Yannick Welsch <[email protected]>
Co-authored-by: Tanguy Leroux <[email protected]>
original-brownbear
added a commit
that referenced
this issue
Oct 23, 2019
This change adds a new field `"shards"` to `RepositoryData` that contains a mapping of `IndexId` to a `String[]`. This string array can be accessed by shard id to get the generation of a shard's shard folder (i.e. the `N` in the name of the currently valid `/indices/${indexId}/${shardId}/index-${N}` for the shard in question).
This allows for creating a new snapshot in the shard without doing any LIST operations on the shard's folder. In the case of AWS S3, this saves about 1/3 of the cost for updating an empty shard (see #45736) and removes one out of two remaining potential issues with eventually consistent blob stores (see #38941 ... now only the root `index-${N}` is determined by listing).
Also and equally if not more important, a number of possible failure modes on eventually consistent blob stores like AWS S3 are eliminated by moving all delete operations to the `master` node and moving from incremental naming of shard level index-N to uuid suffixes for these blobs.
This change moves the deleting of the previous shard level `index-${uuid}` blob to the master node instead of the data node allowing for a safe and consistent update of the shard's generation in the `RepositoryData` by first updating `RepositoryData` and then deleting the now unreferenced `index-${newUUID}` blob.
__No deletes are executed on the data nodes at all for any operation with this change.__
Note also: Previous issues with hanging data nodes interfering with master nodes are completely impossible, even on S3 (see next section for details).
This change changes the naming of the shard level `index-${N}` blobs to a uuid suffix `index-${UUID}`. The reason for this is the fact that writing a new shard-level `index-` generation blob is not atomic anymore in its effect. Not only does the blob have to be written to have an effect, it must also be referenced by the root level `index-N` (`RepositoryData`) to become an effective part of the snapshot repository.
This leads to a problem if we were to use incrementing names like we did before. If a blob `index-${N+1}` is written but due to the node/network/cluster/... crashes the root level `RepositoryData` has not been updated then a future operation will determine the shard's generation to be `N` and try to write a new `index-${N+1}` to the already existing path. Updates like that are problematic on S3 for consistency reasons, but also create numerous issues when thinking about stuck data nodes.
Previously stuck data nodes that were tasked to write `index-${N+1}` but got stuck and tried to do so after some other node had already written `index-${N+1}` were prevented form doing so (except for on S3) by us not allowing overwrites for that blob and thus no corruption could occur.
Were we to continue using incrementing names, we could not do this. The stuck node scenario would either allow for overwriting the `N+1` generation or force us to continue using a `LIST` operation to figure out the next `N` (which would make this change pointless).
With uuid naming and moving all deletes to `master` this becomes a non-issue. Data nodes write updated shard generation `index-${uuid}` and `master` makes those `index-${uuid}` part of the `RepositoryData` that it deems correct and cleans up all those `index-` that are unused.
Co-authored-by: Yannick Welsch <[email protected]>
Co-authored-by: Tanguy Leroux <[email protected]>
This was referenced Feb 3, 2020
rjernst
added
the
Team:Distributed (Obsolete)
original-brownbear
added a commit
that referenced
this issue
Jun 5, 2020
This PR introduces two new fields in to `RepositoryData` (index-N) to track the blob name of `IndexMetaData` blobs and their content via setting generations and uuids. This is used to deduplicate the `IndexMetaData` blobs (`meta-{uuid}.dat` in the indices folders under `/indices` so that new metadata for an index is only written to the repository during a snapshot if that same metadata can't be found in another snapshot.
This saves one write per index in the common case of unchanged metadata thus saving cost and making snapshot finalization drastically faster if many indices are being snapshotted at the same time.
The implementation is mostly analogous to that for shard generations in #46250 and piggy backs on the BwC mechanism introduced in that PR (which means this PR needs adjustments if it doesn't go into `7.6`).
Relates to #45736 as it improves the efficiency of snapshotting unchanged indices
Relates to #49800 as it has the potential of loading the index metadata for multiple snapshots of the same index concurrently much more efficient speeding up future concurrent snapshot delete
original-brownbear
added a commit
to original-brownbear/elasticsearch
that referenced
this issue
Jul 14, 2020
This PR introduces two new fields in to `RepositoryData` (index-N) to track the blob name of `IndexMetaData` blobs and their content via setting generations and uuids. This is used to deduplicate the `IndexMetaData` blobs (`meta-{uuid}.dat` in the indices folders under `/indices` so that new metadata for an index is only written to the repository during a snapshot if that same metadata can't be found in another snapshot.
This saves one write per index in the common case of unchanged metadata thus saving cost and making snapshot finalization drastically faster if many indices are being snapshotted at the same time.
The implementation is mostly analogous to that for shard generations in elastic#46250 and piggy backs on the BwC mechanism introduced in that PR (which means this PR needs adjustments if it doesn't go into `7.6`).
Relates to elastic#45736 as it improves the efficiency of snapshotting unchanged indices
Relates to elastic#49800 as it has the potential of loading the index metadata for multiple snapshots of the same index concurrently much more efficient speeding up future concurrent snapshot delete
original-brownbear
added a commit
that referenced
this issue
Jul 14, 2020
This PR introduces two new fields in to `RepositoryData` (index-N) to track the blob name of `IndexMetaData` blobs and their content via setting generations and uuids. This is used to deduplicate the `IndexMetaData` blobs (`meta-{uuid}.dat` in the indices folders under `/indices` so that new metadata for an index is only written to the repository during a snapshot if that same metadata can't be found in another snapshot.
This saves one write per index in the common case of unchanged metadata thus saving cost and making snapshot finalization drastically faster if many indices are being snapshotted at the same time.
The implementation is mostly analogous to that for shard generations in #46250 and piggy backs on the BwC mechanism introduced in that PR (which means this PR needs adjustments if it doesn't go into `7.6`).
Relates to #45736 as it improves the efficiency of snapshotting unchanged indices
Relates to #49800 as it has the potential of loading the index metadata for multiple snapshots of the same index concurrently much more efficient speeding up future concurrent snapshot delete
|
Is this still relevant? |
|
@ywelsch IMO it is. We still write two blobs for every unchanged shard snapshot despite all the fixes made. Finding a way for not doing that would still be a worthwhile thing to do whenever there is some time for it. |
|
Is it right that we're now down to one blob per unchanged shard snapshot, namely the I mean this is still an issue but it's half of the issue it used to be. Fixing it would require some more indirection somewhere - I guess we could record the shard-level ID in |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Currently, a snapshot of a shard that has not changed at all relative to an existing snapshot of the shard (i.e. does not require uploading any files for that shard) still triggers the following operations:
snap-${uuid}.datblob to the unchanged shard's folder in the repositoryindex-Nblob to the unchanged shard's folder in the repositoryIn practice the effect of this is significant for use cases like rolling indices per day/hour/etc. A cluster that contains a small and bounded number of indices/shards that are actively written to and a large and growing number of shards that are constant in time will over time see ever more expensive and slower snapshots even though the amount of data added by each snapshot is not increasing.
This could be avoided by referencing the content of
snap-${uuid}.datin each shard differently. Instead of creating a blob per snapshot+shard tuple, a certain state of a shard could be described by what is currently asnap-${uuid}.datand then itself be referenced from the root levelindex-Nin the repository.The text was updated successfully, but these errors were encountered: