dombizita commented on code in PR #560: URL: https://github.com/apache/ozone-site/pull/560#discussion_r4452841518
########## docs/08-developer-guide/01-develop/05-upgrade-compatibility.md: ########## @@ -0,0 +1,359 @@ +--- +sidebar_label: Upgrade Compatibility +--- + +# Developing for Compatibility with Zero Downtime Upgrade + +Zero Downtime Upgrade (ZDU) lets the internal components of an Ozone cluster (OM, SCM, Datanode, S3 Gateway, and Recon) run in mixed versions during an upgrade with no interruption to service. Additionally, Ozone provides a window where a cluster can be downgraded after the upgrade. Both of these features require every release to uphold certain compatibility guarantees; otherwise, an upgrade can silently introduce problems such as diverging OM state machines, a corrupted container replica, or a Datanode that is unable to re-register. This page is a practical guide to what counts as an incompatible change inside the cluster, and how to safely introduce such changes in each area of the code. + +The rationale behind the framework lives in the [ZDU design document](https://github.com/apache/ozone/blob/master/hadoop-hdds/docs/content/design/zdu-design.md). Read it if you want the full model. This page focuses only on day-to-day development. + +:::note Scope +This page covers compatibility **between internal Ozone components** during an upgrade. External client ↔ Ozone-server compatibility is a separate concern. +::: + +## Background + +### Vocabulary + +- **Apparent version**: The version a component is *acting as*. It is persisted to disk and determines both the API the component exposes and the on-disk format it writes. + +- **Software version**: The highest component version contained in the running code. This is fixed by the bits that are installed. + +- **Pre-finalized**: The state a component is in when its apparent version on disk is *less than* its software version. All old features work; new features are blocked; downgrade is allowed. Components enter this state when they are first started after an upgrade, and remain here until they are either downgraded or finalized by the admin. + +- **Finalized**: The state a component reaches when its apparent version *equals* its software version. All features are allowed; downgrade is no longer possible. Components enter this state after they have been upgraded to a pre-finalized state and then given an explicit finalize command by the admin. This transition happens while the cluster is running and the cluster is expected to remain operational during finalization. + +Any behavior incompatible with an old version's disk layout or API must be mapped to a version and disabled until its corresponding version is finalized. + +### Invariants Provided to the Developer + +To maintain consistency during zero downtime upgrade, the following invariants can be relied on by developers: + +1. **For internal client/server relationships, the server is always upgraded before its client.** + - The only exception is Recon, which is a client of OM but is upgraded ahead of it. + - This is maintained by the admin following the documented component upgrade order. +2. **All internal components are running the newest software before finalization begins.** + - Ozone's upgrade framework provides best-effort enforcement of this invariant, but due to corner cases it ultimately falls on the admin to execute correctly. +3. **For internal client/server relationships, the server is always finalized before its client.** + - Again, Recon is the only exception, since it finalizes before OM. + - This is fully enforced by the upgrade framework within Ozone. + +#### Upgrade Order + +Components are upgraded in the following order to meet the client/server version invariant: + +- SCM +- Recon +- Datanodes +- OM +- S3 and HttpFS Gateways (can be upgraded simultaneously) + +Feature developers do not need to account for upgrade orders that deviate from this when handling compatibility for a feature. An out-of-order upgrade is considered user error. + +#### Finalization Order + +Components finalize in the same order they are upgraded. Finalization is triggered by a single admin command and the cluster ensures the correct finalization order internally. + +- SCM: Finalized via Ratis. +- Datanodes: Finalized asynchronously by SCM after it finalizes itself. +- OM: Finalized via Ratis only once HDDS (SCM and all live Datanodes) have finalized. + +Stateless gateway components do not need to finalize since they have no disk state to manage on downgrade and do not communicate with each other. Their compatibility with the rest of the cluster is handled solely by the guarantee that the server they call remains backwards compatible with older client versions. Recon currently finalizes immediately on startup, although downgrade support for Recon may be added in the future. + +## Quick Start + +Within this framework, most incompatible changes can be handled by upholding two rules during feature development: + +1. **Components with the same apparent version expose the same API surface and persist data in the same format.** +2. **For internal client/server relationships, the server remains backwards compatible with older clients.** + +When a change to API surface or disk layout must be introduced, a new component version must be added to gate the feature. To do this, first add an entry to the affected components' version enums: + +- [`OzoneManagerVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-hdds/common/src/main/java/org/apache/hadoop/ozone/OzoneManagerVersion.java) — used within the OM Ratis group and reported to external clients. +- [`HDDSVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-hdds/common/src/main/java/org/apache/hadoop/hdds/HDDSVersion.java) — shared by SCM and Datanodes so SCM can orchestrate Datanode finalization. It is also shared with clients. +- [`ReconVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-ozone/recon/src/main/java/org/apache/hadoop/ozone/recon/upgrade/ReconVersion.java) — Recon's own internal disk-format version, which is currently not shared with other components. + +:::warning +Do **not** add new entries to `OMLayoutFeature` or `HDDSLayoutFeature`. Those enums are frozen after the migration to ZDU's new versioning framework. Every new version goes into `OzoneManagerVersion`, `HDDSVersion`, or `ReconVersion`. +::: + +```java +public enum OzoneManagerVersion implements ComponentVersion { + // Previous versions... + MY_NEW_FEATURE(101 /*or next available version integer*/, "My new feature"), +} +``` + +Most incompatible changes are handled by gating the new behavior on the component's apparent version through its version manager: + +```java +if (ozoneManager.getVersionManager().isAllowed(OzoneManagerVersion.MY_NEW_FEATURE)) { + // All capabilities of MY_NEW_FEATURE can be used, even if they are incompatible with older components. +} else { + // Must continue to act as the older version without using incompatible changes from MY_NEW_FEATURE. +} +``` + +`isAllowed(...)` returns true only once the component's apparent version is at least the version you name. Each component has a version manager with an equivalent method. + +:::warning +It is incorrect to write the following: + +```java +if (!ozoneManager.getVersionManager().needsFinalization()) { + // Use MY_NEW_FEATURE +} else { + // Bypass MY_NEW_FEATURE +} +``` + +New versions will continue to be added in subsequent releases after `MY_NEW_FEATURE`. Say `MY_NEW_FEATURE_2` is added in the next release. `needsFinalization()` is true whenever the component is pre-finalized for any newer version. A cluster that is pre-finalized for `MY_NEW_FEATURE_2` would therefore wrongly disable `MY_NEW_FEATURE`, even though that feature is already finalized. +::: + +## Committing Changes + +Ozone's component versions are a logical versioning framework decoupled from official Ozone releases. The system is expected to work correctly regardless of which commit a release is cut from. To maintain this property, changes requiring versioning must be checked in within the same commit that introduces the version. Splitting this across multiple commits will create an incompatibility if a release is cut between the commits: + +1. **Incompatible: Adding an incompatible change and then adding a version gate for it in a follow-up commit.** A release cut between these two commits will release the incompatible change without the version gate. + +2. **Incompatible: Adding a version in one commit and using it in a follow-up commit.** Clustrers on a release cut between these two commits will finalize the version without the feature code. The next release's code then activates the feature with no version gate or upgrade action. Review Comment: ```suggestion 2. **Incompatible: Adding a version in one commit and using it in a follow-up commit.** Clusters on a release cut between these two commits will finalize the version without the feature code. The next release's code then activates the feature with no version gate or upgrade action. ``` ########## docs/08-developer-guide/01-develop/05-upgrade-compatibility.md: ########## @@ -0,0 +1,359 @@ +--- +sidebar_label: Upgrade Compatibility +--- + +# Developing for Compatibility with Zero Downtime Upgrade + +Zero Downtime Upgrade (ZDU) lets the internal components of an Ozone cluster (OM, SCM, Datanode, S3 Gateway, and Recon) run in mixed versions during an upgrade with no interruption to service. Additionally, Ozone provides a window where a cluster can be downgraded after the upgrade. Both of these features require every release to uphold certain compatibility guarantees; otherwise, an upgrade can silently introduce problems such as diverging OM state machines, a corrupted container replica, or a Datanode that is unable to re-register. This page is a practical guide to what counts as an incompatible change inside the cluster, and how to safely introduce such changes in each area of the code. + +The rationale behind the framework lives in the [ZDU design document](https://github.com/apache/ozone/blob/master/hadoop-hdds/docs/content/design/zdu-design.md). Read it if you want the full model. This page focuses only on day-to-day development. + +:::note Scope +This page covers compatibility **between internal Ozone components** during an upgrade. External client ↔ Ozone-server compatibility is a separate concern. +::: + +## Background + +### Vocabulary + +- **Apparent version**: The version a component is *acting as*. It is persisted to disk and determines both the API the component exposes and the on-disk format it writes. + +- **Software version**: The highest component version contained in the running code. This is fixed by the bits that are installed. + +- **Pre-finalized**: The state a component is in when its apparent version on disk is *less than* its software version. All old features work; new features are blocked; downgrade is allowed. Components enter this state when they are first started after an upgrade, and remain here until they are either downgraded or finalized by the admin. + +- **Finalized**: The state a component reaches when its apparent version *equals* its software version. All features are allowed; downgrade is no longer possible. Components enter this state after they have been upgraded to a pre-finalized state and then given an explicit finalize command by the admin. This transition happens while the cluster is running and the cluster is expected to remain operational during finalization. + +Any behavior incompatible with an old version's disk layout or API must be mapped to a version and disabled until its corresponding version is finalized. + +### Invariants Provided to the Developer + +To maintain consistency during zero downtime upgrade, the following invariants can be relied on by developers: + +1. **For internal client/server relationships, the server is always upgraded before its client.** + - The only exception is Recon, which is a client of OM but is upgraded ahead of it. + - This is maintained by the admin following the documented component upgrade order. +2. **All internal components are running the newest software before finalization begins.** + - Ozone's upgrade framework provides best-effort enforcement of this invariant, but due to corner cases it ultimately falls on the admin to execute correctly. +3. **For internal client/server relationships, the server is always finalized before its client.** + - Again, Recon is the only exception, since it finalizes before OM. + - This is fully enforced by the upgrade framework within Ozone. + +#### Upgrade Order + +Components are upgraded in the following order to meet the client/server version invariant: + +- SCM +- Recon +- Datanodes +- OM +- S3 and HttpFS Gateways (can be upgraded simultaneously) + +Feature developers do not need to account for upgrade orders that deviate from this when handling compatibility for a feature. An out-of-order upgrade is considered user error. + +#### Finalization Order + +Components finalize in the same order they are upgraded. Finalization is triggered by a single admin command and the cluster ensures the correct finalization order internally. + +- SCM: Finalized via Ratis. +- Datanodes: Finalized asynchronously by SCM after it finalizes itself. +- OM: Finalized via Ratis only once HDDS (SCM and all live Datanodes) have finalized. + +Stateless gateway components do not need to finalize since they have no disk state to manage on downgrade and do not communicate with each other. Their compatibility with the rest of the cluster is handled solely by the guarantee that the server they call remains backwards compatible with older client versions. Recon currently finalizes immediately on startup, although downgrade support for Recon may be added in the future. + +## Quick Start + +Within this framework, most incompatible changes can be handled by upholding two rules during feature development: + +1. **Components with the same apparent version expose the same API surface and persist data in the same format.** +2. **For internal client/server relationships, the server remains backwards compatible with older clients.** + +When a change to API surface or disk layout must be introduced, a new component version must be added to gate the feature. To do this, first add an entry to the affected components' version enums: + +- [`OzoneManagerVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-hdds/common/src/main/java/org/apache/hadoop/ozone/OzoneManagerVersion.java) — used within the OM Ratis group and reported to external clients. +- [`HDDSVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-hdds/common/src/main/java/org/apache/hadoop/hdds/HDDSVersion.java) — shared by SCM and Datanodes so SCM can orchestrate Datanode finalization. It is also shared with clients. +- [`ReconVersion.java`](https://github.com/apache/ozone/blob/master/hadoop-ozone/recon/src/main/java/org/apache/hadoop/ozone/recon/upgrade/ReconVersion.java) — Recon's own internal disk-format version, which is currently not shared with other components. + +:::warning +Do **not** add new entries to `OMLayoutFeature` or `HDDSLayoutFeature`. Those enums are frozen after the migration to ZDU's new versioning framework. Every new version goes into `OzoneManagerVersion`, `HDDSVersion`, or `ReconVersion`. +::: + +```java +public enum OzoneManagerVersion implements ComponentVersion { + // Previous versions... + MY_NEW_FEATURE(101 /*or next available version integer*/, "My new feature"), +} +``` + +Most incompatible changes are handled by gating the new behavior on the component's apparent version through its version manager: + +```java +if (ozoneManager.getVersionManager().isAllowed(OzoneManagerVersion.MY_NEW_FEATURE)) { + // All capabilities of MY_NEW_FEATURE can be used, even if they are incompatible with older components. +} else { + // Must continue to act as the older version without using incompatible changes from MY_NEW_FEATURE. +} +``` + +`isAllowed(...)` returns true only once the component's apparent version is at least the version you name. Each component has a version manager with an equivalent method. + +:::warning +It is incorrect to write the following: + +```java +if (!ozoneManager.getVersionManager().needsFinalization()) { + // Use MY_NEW_FEATURE +} else { + // Bypass MY_NEW_FEATURE +} +``` + +New versions will continue to be added in subsequent releases after `MY_NEW_FEATURE`. Say `MY_NEW_FEATURE_2` is added in the next release. `needsFinalization()` is true whenever the component is pre-finalized for any newer version. A cluster that is pre-finalized for `MY_NEW_FEATURE_2` would therefore wrongly disable `MY_NEW_FEATURE`, even though that feature is already finalized. +::: + +## Committing Changes + +Ozone's component versions are a logical versioning framework decoupled from official Ozone releases. The system is expected to work correctly regardless of which commit a release is cut from. To maintain this property, changes requiring versioning must be checked in within the same commit that introduces the version. Splitting this across multiple commits will create an incompatibility if a release is cut between the commits: + +1. **Incompatible: Adding an incompatible change and then adding a version gate for it in a follow-up commit.** A release cut between these two commits will release the incompatible change without the version gate. + +2. **Incompatible: Adding a version in one commit and using it in a follow-up commit.** Clustrers on a release cut between these two commits will finalize the version without the feature code. The next release's code then activates the feature with no version gate or upgrade action. + +3. **Incompatible: Extending the usage of an existing version in follow-up commits.** This breaks for the same reason as (2). A new version must be used for the new functionality. + +One large feature may introduce incompatible changes in multiple areas which take multiple commits to add; however, it is easier to reason about these if they are grouped under one component version. There are two strategies developers can use for this: + +1. **Build these features on feature branches.** Feature branches are not release candidates, so they can contain incompatible changes between commits. A single version just needs to be added to gate all incompatible behavior before the branch is merged back to master. + +2. **Place the incompatible changes behind a disabled feature flag, and switch this flag to a version gate when the feature is ready to be released.** If multiple incompatible changes for a feature are being added to master, they can each be gated with a feature flag while under development. If the feature flag must be exposed as a configuration key, ensure it has a prefix (like `dev.`) and a description that indicates it enables an unsupported feature and should only be used for development. While developing the feature, use this pattern: + +```java +// featureFlagEnabled = false except when needed in test environments. +if (featureFlagEnabled) { + // All capabilities of MY_NEW_FEATURE can be used, even if they are incompatible with older components. +} else { + // Must continue to act as the older version without using incompatible changes from MY_NEW_FEATURE. +} +``` + +Once all incompatible changes are added over multiple commits, replace all usages of the feature flag with a version gate in a single commit to release the feature. This makes it identical to the pattern shown above: + +```diff +--- a/example.java ++++ b/example.java +@@ -1,6 +1,5 @@ +-// featureFlagEnabled = false except when needed in test environments. +-if (featureFlagEnabled) { ++if (ozoneManager.getVersionManager().isAllowed(OzoneManagerVersion.MY_NEW_FEATURE)) { + // All capabilities of MY_NEW_FEATURE can be used, even if they are incompatible with older components. + } else { + // Must continue to act as the older version without using incompatible changes from MY_NEW_FEATURE. + } +``` + +:::warning +Developing large features on master must be done with care. The next section documents many surfaces that are susceptible to incompatible changes, and large features have the potential to accidentally break these if a release is cut from master while the feature is still being developed. Since protobuf structures and server-side APIs cannot accept incompatible changes after a release, only features built on feature branches have the freedom to create initial draft implementations of these and update them incompatibly as the design evolves. +::: + +## Surface-by-Surface Guide + +Most surfaces below follow the same shape: **what is considered incompatible** and **the pattern to handle it**. Note that OM and SCM are finalized via Ratis, but Datanodes finalize asynchronously as they receive the finalize command from SCM. + +### General Protobuf Compatibility + +Most of Ozone's disk and network serialization formats are protobuf. Basic protobuf compatibility is statically enforced by Ozone's CI using [protolock](https://github.com/nilslice/protolock). However, it is possible to break compatibility at the proto layer while still passing protolock. This usually happens when the component writing the protobuf is newer than the component reading it. A new writer/old reader scenario may occur when a proto is written to disk in a new version before finalization and read by an old version after downgrade. It may also occur in internal client/server relationships when the server is upgraded and returns a response to an older client. The following changes may still need additional version gating for upgrades despite being compatible with protolock: + +**Potentially Incompatible - New Fields**: Old readers will silently ignore a field they do not know about. Old writers will leave the new field empty and new readers will fall back to the default value. + +**Pattern**: If no versioning is used, the field's default value has to be the safe interpretation. The writer must behave acceptably when the field is ignored, and the reader must behave acceptably when the field is missing. Otherwise, a corresponding component version must be added and the field cannot be used until that version is finalized. + +**Incompatible - New Enum Values**: Even though protobuf can carry an unrecognized enum value over the wire, Ozone code switches on these values to decide what to do, so an old reader has no correct branch to take. + +**Pattern**: New enum values require version gating and must not be written until their corresponding version is finalized. + +### Configuration Changes + +**Incompatible:** Any change to an existing configuration's default value, or a new configuration with a default value that impacts: + +- Communication between components with and without the configuration +- Downgrade after the configuration is used + +During a rolling upgrade, older clients without the configuration must still be able to communicate with newer servers that have the configuration. Additionally, before finalization, all disk state must remain readable by the old version regardless of any supplied configuration values. + +**Pattern:** A configuration that may interfere with downgrade should not be acted on until after its corresponding component version is finalized. If the configuration is only relevant at startup, it will be re-read the next time the component is restarted after finalization completes. Configuration changes that impact network communication between components cannot happen automatically during ZDU. Applying them during a rolling restart outside of the upgrade would produce the same problem, so such changes must be made with partial downtime on a case-by-case basis. + +### New OM Requests (Read or Write) + +**Incompatible:** Any new read or write request added to the OM. New write requests cannot be applied by all nodes when the cluster is in a mixed-version state. New read requests may appear and disappear to the client as a mixed-version OM Ratis group changes leaders. + +**Pattern:** Block new write requests until their feature is finalized by putting `@DisallowedUntilOmVersion(OzoneManagerVersion.X)` on the `preExecute` method of its `OMClientRequest` subclass. Block new read requests until their feature is finalized by putting `@DisallowedUntilOmVersion(OzoneManagerVersion.X)` on the helper method that the read delegates to in `OzoneManagerRequestHandler`. + +### New Fields or Behavior in Existing OM Requests (Read or Write) + +**Incompatible:** Any new option that can be passed in by a client that changes the result of an existing read or write request. For example, adding a new replication type that can be requested in the `allocateBlock` request. + +**Pattern:** Intercept the request before the leader OM processes it using a `RequestFeatureValidator`. Check if the incompatible argument has been specified, and if so, fail the request. This is typically placed in the request's `OMClientRequest` subclass for write requests and in `OzoneManagerRequestHandler` for read requests. + +This is a real example of using `RequestFeatureValidator` to block a new incompatible field in an existing request until the corresponding version is finalized. Note that this uses the old `OMLayoutFeature` version enum instead of the current `OzoneManagerVersion` enum. + +```java + @RequestFeatureValidator( + conditions = ValidationCondition.CLUSTER_NEEDS_FINALIZATION, + processingPhase = RequestProcessingPhase.PRE_PROCESS, + requestType = Type.SetBucketProperty + ) + public static OMRequest disallowSetBucketPropertyWithECReplicationConfig( + OMRequest req, ValidationContext ctx) throws OMException { + if (!ctx.versionManager().isAllowed(OMLayoutFeature.ERASURE_CODED_STORAGE_SUPPORT)) { + SetBucketPropertyRequest propReq = req.getSetBucketPropertyRequest(); + if (propReq.hasBucketArgs() + && propReq.getBucketArgs().hasDefaultReplicationConfig() + && propReq.getBucketArgs().getDefaultReplicationConfig() + .hasEcReplicationConfig()) { + throw new OMException("Cluster does not have the Erasure Coded" + + " Storage support feature finalized yet, but the request contains" + + " an Erasure Coded replication type. Rejecting the request," + + " please finalize the cluster upgrade and then try again.", + OMException.ResultCodes.NOT_SUPPORTED_OPERATION_PRIOR_FINALIZATION); + } + } + return req; + } +``` + +### OM Apply Transaction / OM RocksDB State + +**Incompatible:** Any change to what an OM request writes to RocksDB inside `validateAndUpdateCache`. This includes: + +- A new column family +- A change to the shape of a persisted value class +- Addition or removal of exceptions that abort further processing of the request + - The version without the exception will persist the update to the state machine, while the version with the exception will abort without updating the state machine. +- A difference in how a persisted value is calculated, even if the protobuf schema does not change. + - For example, an update to quota calculation logic would require version gating. + +**Pattern:** Gate the new persisted behavior in `validateAndUpdateCache` on `ozoneManager.getVersionManager().isAllowed(OzoneManagerVersion.X)` and keep the pre-feature code path for the negative branch. Two OMs can replay the same Ratis transaction while acting as different versions, so the gate keeps their state machines from diverging. Since the state machines remain compatible with older versions, this also allows mixed-version leader/follower setups to process the contents of Ratis snapshots. + +If in doubt, ask: *would an OM acting as apparent version X and an OM acting as apparent version Y, both replaying this transaction from Ratis, arrive at the same DB state?* If not, it needs a gate. + +### OM/SCM ↔ OM/SCM Peer RPCs Outside Ratis, Including Ratis Snapshot Installation APIs + +**Incompatible:** Adding a required method, or changing the semantics of an existing method, in [`OMAdminProtocol`](https://github.com/apache/ozone/blob/master/hadoop-ozone/common/src/main/java/org/apache/hadoop/ozone/om/protocol/OMAdminProtocol.java) or [`OMInterServiceProtocol`](https://github.com/apache/ozone/blob/master/hadoop-ozone/common/src/main/java/org/apache/hadoop/ozone/om/protocol/OMInterServiceProtocol.java). During a rolling upgrade, peer OMs can be at mixed software versions, so a new method may not exist on the callee. Unlike an unknown protobuf *field*, an unknown *method* fails the RPC outright. Ratis snapshots are also installed user a peer to peer API, so they are subject to the same requirements. Review Comment: ```suggestion **Incompatible:** Adding a required method, or changing the semantics of an existing method, in [`OMAdminProtocol`](https://github.com/apache/ozone/blob/master/hadoop-ozone/common/src/main/java/org/apache/hadoop/ozone/om/protocol/OMAdminProtocol.java) or [`OMInterServiceProtocol`](https://github.com/apache/ozone/blob/master/hadoop-ozone/common/src/main/java/org/apache/hadoop/ozone/om/protocol/OMInterServiceProtocol.java). During a rolling upgrade, peer OMs can be at mixed software versions, so a new method may not exist on the callee. Unlike an unknown protobuf *field*, an unknown *method* fails the RPC outright. Ratis snapshots are also installed using a peer to peer API, so they are subject to the same requirements. ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
