[
https://issues.apache.org/jira/browse/FLINK-41435?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
LuChengxiang updated FLINK-41435:
---------------------------------
Description:
h2. Summary
Flink periodically obtains HDFS delegation tokens and distributes them through
HadoopDelegationTokenReceiver. The receiver adds refreshed credentials to
UserGroupInformation.getCurrentUser() of the thread running the refresh. A
long-lived Hadoop client, especially a checkpoint-storage DFSClient, may
instead use the UGI/Subject captured when that client was created. If those
identities differ, the token refresh succeeds and the receiver reports success,
but the new credentials are not delivered to the client's UGI. The client
continues using its startup token and fails at the 24h HDFS token deadline with
SecretManager$InvalidToken.
In the observed Kubernetes application-mode reproduction, the mismatch becomes
reachable through ForkJoinPool's managed-block compensation. The provider
blocks in {{CompletableFuture.get()}} on a ForkJoinWorkerThread; when no idle
FJP thread is available and the pool has capacity, the compensation path
creates additional dispatcher threads while {{freshUGI.doAs(...)}} is
effective, and those threads inherit the fresh Subject. A later {{cluster-io}}
thread created by one of them inherits the same Subject and can run a token
receiver under it.
The defect is receiver target selection, not token acquisition or client
binding. The proposed fix targets the UGI held by HadoopSecurityContext,
falling back to the login UGI, instead of a pooled thread's current UGI. Here
"UGI" means the Hadoop UserGroupInformation wrapper and "Subject" means the
identity it carries; identity comparisons use Subject identity.
h2. Observed failure
This was observed in a Flink 1.20.5 Kubernetes application-mode job using
ZooKeeper-based high availability and HDFS-backed checkpoint storage. The
primary failure was in the checkpoint storage path:
{noformat}
CheckpointCoordinator.initializeCheckpointLocation
-> FsCheckpointStorageAccess.initializeLocationForCheckpoint
-> HadoopFileSystem.mkdirs
-> DFSClient.mkdirs
{noformat}
After the 24h deadline, the checkpoint-storage client reported InvalidToken for
its startup token even though refresh had obtained a new HDFS token and the
receiver reported success.
Not every Kubernetes application-mode job is affected. The failure is
timing-dependent: a long-lived Hadoop client must capture one Subject/UGI
lifecycle, and a later refresh receiver must run under a different one.
h2. Root cause
The delivery bug is in HadoopDelegationTokenReceiver. It targets
{{UserGroupInformation.getCurrentUser()}} of the refresh thread.
{{getCurrentUser()}} resolves from the thread's effective AccessControlContext
and is not guaranteed to be the UGI held by HadoopSecurityContext or the UGI
captured by a long-lived checkpoint-storage DFSClient. In the observed failure,
the checkpoint client's captured UGI had the login/HadoopSecurityContext
Subject, while the receiver could run under a pooled thread's inherited fresh
Subject.
h3. Acquisition and refresh phases
Token delivery in the JobManager has three phases. Within each phase, the
provider and the JM-local receiver run on the same token-manager thread.
* Bootstrap acquisition: {{ClusterEntrypoint.initializeServices()}} calls
{{DefaultDelegationTokenManager.obtainDelegationTokens()}} on {{main}} before
BlobServer starts. It updates local receivers for early external filesystem
access. The receiver runs after {{freshUGI.doAs(...)}} returns, when current is
login, so this phase is safe.
* ResourceManager startup acquisition: {{ResourceManager.onStart()}} ->
{{DefaultDelegationTokenManager.start()}} -> {{startTokensUpdate()}} runs on
{{default-dispatcher}}. This provider call is where the observed ForkJoinPool
expansion occurs, but the startup receiver itself still runs on
{{default-dispatcher}} after the provider returns, so it does not target the
fresh Subject.
* Scheduled refresh: later updates are submitted through {{cluster-io}}. A
receiver running on a {{cluster-io}} thread that inherited the fresh Subject
resolves {{getCurrentUser()}} to that Subject and can target the wrong UGI.
TaskManager receivers are invoked by {{TaskExecutor}} token updates, not on
{{cluster-io}}; the probe in this issue covers the JM-local receiver path only.
These two startup token acquisitions are separate from the double
{{SecurityUtils.install}} sequence discussed under Scope and boundaries.
h3. How the fresh Subject reaches a pooled thread
The observed Subject-propagation window is the ResourceManager startup
acquisition. The provider runs on a ForkJoinWorkerThread. The fresh Subject
comes from {{KerberosLoginProvider.doLoginAndReturnUGI()}}, which calls
{{UserGroupInformation.loginUserFromKeytabAndReturnUGI(principal, keytab)}}.
That creates a new UGI for the same principal and does not replace the process
login user.
While {{freshUGI.doAs(...)}} is effective, the provider action calls
{{fs.addDelegationTokens(...)}}; the blocking HDFS RPC in
{{CompletableFuture.get()}} enters ForkJoinPool's managed-block path. If no
idle FJP thread is available and the pool has capacity, {{tryCompensate()}}
creates additional dispatcher threads during the blocked RPC. This compensation
does not unblock the RPC; it preserves pool capacity for other work. The
{{Thread}} constructor for each new thread captures the effective
AccessControlContext returned by {{AccessController.getContext()}} into that
thread's {{inheritedAccessControlContext}}; this is a stack-derived snapshot,
not the entire live Java stack, and it includes the Subject active during
{{doAs}}. After {{doAs}} returns, current on the provider thread resolves to
login again, but the newly created threads retain the fresh Subject.
The probe block below combines the observed call chain, Subject identities, and
thread creation. Here HadoopSecurityContext Subject means the Subject of the
UGI held by the installed HadoopSecurityContext; it matches the login Subject
in this probe, but that is not guaranteed in general.
{noformat}
provider thread = default-dispatcher-12
freshUGI Subject = 817147333
login/HadoopSecurityContext Subject = 1333938290
freshUGI.doAs(...)
-> fs.addDelegationTokens(...)
-> HDFS RPC blocked in CompletableFuture.get()
-> ForkJoinPool.managedBlock()
-> tryCompensate()
-> createWorker() [when no idle FJP thread and pool capacity remains]
-> new dispatcher threads inherit the fresh Subject
-> later, one of them triggers creation of a cluster-io thread
current Subject before doAs = 1333938290
current Subject inside doAs = 817147333
current Subject after doAs = 1333938290
FJP-created dispatcher threads:
default-dispatcher-13 / 14 / 15, creator=default-dispatcher-12
later cluster-io thread:
cluster-io-thread-3, creator=default-dispatcher-15
current Subject / JM-local receiver target (getCurrentUser()) = 817147333
login/HadoopSecurityContext Subject = 1333938290
{noformat}
{{fs.addDelegationTokens(...)}} is the provider-side acquisition step: it
obtains the filesystem delegation token or tokens and stores them in a local
{{Credentials}} object. The receiver later delivers that object to the selected
UGI.
{{cluster-io}} is a fixed-size executor; its threads are created lazily on task
submission. In the probe, an FJP-created dispatcher thread later submits a task
to this executor, which creates a cluster-io thread that copies the submitter's
inherited ACC and starts with the fresh Subject. This thread creation is not
the token refresh itself; it only determines the Subject under which the later
scheduled refresh runs. A receiver invocation on that thread resolves
{{getCurrentUser()}} to the fresh Subject and writes credentials there, while
the login and HadoopSecurityContext UGIs remain unchanged and still resolve to
the login Subject shown above.
This is a per-thread inherited security context, not a global mutation of
UserGroupInformation. It is a receiver target-selection mismatch, not a
client-binding error. Clients created under an unrelated transient Subject
require separate rebinding or recreation and are out of scope.
h2. Proposed change
This issue fixes token delivery for the observed long-lived-client case,
primarily checkpoint storage.
* Add an accessor for the UGI held by HadoopSecurityContext.
* In HadoopDelegationTokenReceiver, select a single target UGI: use the UGI
held by HadoopSecurityContext when present; otherwise fall back to
{{UserGroupInformation.getLoginUser()}}. Add refreshed credentials only to that
UGI. Do not target {{UserGroupInformation.getCurrentUser()}}, which may be a
pooled thread's inherited Subject.
* Add unit tests for the receiver's context-first and login-fallback target
selection.
The change makes the delivery target deterministic. It does not rebind a client
already created under a different Subject.
h2. Scope and boundaries
h3. Out of scope
* Preventing HDFS client initialization or blocking RPCs from occurring inside
{{freshUGI.doAs(...)}} or the provider's ForkJoinPool thread.
* Closing or rebinding FileSystem instances created under a transient fresh
Subject; this is a separate client-lifecycle cleanup.
* Delivering credentials to clients created under an unrelated transient
Subject; those clients require separate rebinding or recreation.
* Changing the provider's {{getCurrentUser()}} usage.
{{HadoopFSDelegationTokenProvider}} already acquires tokens inside
{{KerberosLoginProvider.doLoginAndReturnUGI()}}; its remaining
{{getCurrentUser()}} use is a token-required pre-check and renewer-name lookup.
No provider change is required for the observed receiver mismatch; it remains a
separate follow-up if provider-side behavior is observed.
* Double security installation and {{SecurityUtils.install}} idempotency. A
diagnostic build where the second installation was skipped still failed at the
24h deadline. The early/duplicate installation adds lifecycle complexity, but
the probe does not establish it as a cause of the receiver mismatch.
h3. YARN
YARN is not a confirmed reproduction. The YARN samples below were collected
with a Java 8-compatible 1.20.5 probe build; Java 8 remains supported but
deprecated in the 1.20 source line. Across repeated controlled application-mode
probes, some provider windows created no dispatcher thread, while one window
showed a new dispatcher thread during the observation period. However, the FJP
stack probe did not capture a provider-specific
{{tryCompensate}}/{{createWorker}} stack; the observed stacks came from
JobGraph generation and RPC actor paths. The new thread therefore cannot be
attributed to the token provider from the current evidence. One later sample
recorded {{futureDone=false}} immediately before the provider RPC future and
{{true}} after; this shows the future was not yet complete at that point, but
does not prove that {{CompletableFuture.get()}} entered managed blocking,
because it may have completed before the wait. No provider-specific FJP stack
was captured in that sample either. A controlled YARN probe that instrumented
{{CompletableFuture.waitingGet}} and {{ForkJoinPool.managedBlock}} produced no
provider stack, while the same agent produced the expected
waiting-get/managed-block stack in a self-test. That sample therefore did not
enter the provider wait path. The new thread observed in another run remains
unattributed, and the per-window FJP state, timing, and concurrent work remain
unestablished as the explanation.
h3. Versions
Before 1.17, the legacy {{HadoopModule}} path added credentials directly to the
login user; the {{HadoopDelegationTokenReceiver}} /
{{UserGroupInformation.getCurrentUser()}} path starts with 1.17. 1.18, 1.19,
and 1.20 share that receiver behavior. 1.18 is not claimed to be immune.
h2. Testing
* Add receiver unit tests for (a) HadoopSecurityContext UGI precedence with
login/current not updated, (b) login fallback when no HadoopSecurityContext is
installed, and (c) the inherited current UGI not targeted.
* Trigger multiple post-startup token refreshes in a long-running job and
verify which executor runs the provider, whether new ForkJoinPool threads are
created, and which Subject the receiver's current user resolves to.
* Repeat the YARN provider-window probe while instrumenting the provider's
{{CompletableFuture}} state; the current YARN sample is summarized under Scope
and boundaries.
* A deployed application-mode job should survive the 24h token deadline and
continue checkpointing without SecretManager$InvalidToken. Verify that the
checkpoint-storage DFSClient receives the new token.
* Run a controlled application-mode probe that verifies the receiver updates
the UGI selected by the new logic rather than a pooled thread's inherited
current UGI; cover the TaskManager receiver path separately.
h2. Acceptance criteria
* A token refresh updates the UGI held by HadoopSecurityContext when present,
otherwise the login UGI, and does not use a pooled thread's inherited current
UGI; a checkpoint-storage client bound to the selected UGI receives the new
token.
* A long-running application-mode job survives the 24h HDFS token deadline.
was:
h2. Summary
Flink periodically obtains HDFS delegation tokens and distributes them through
HadoopDelegationTokenReceiver. The receiver adds refreshed credentials to
UserGroupInformation.getCurrentUser() of the thread running the refresh. A
long-lived Hadoop client, especially a checkpoint-storage DFSClient, may
instead use the UGI/Subject captured when that client was created. If those
identities differ, the token refresh succeeds and the receiver reports success,
but the new credentials are not delivered to the client's UGI. The client
continues using its startup token and fails at the 24h HDFS token deadline with
SecretManager$InvalidToken.
In the observed Kubernetes application-mode reproduction, the mismatch becomes
reachable through ForkJoinPool's managed-block compensation. The provider
blocks in {{CompletableFuture.get()}} on a ForkJoinWorkerThread; when no idle
FJP thread is available and the pool has capacity, the compensation path
creates additional dispatcher threads while {{freshUGI.doAs(...)}} is
effective, and those threads inherit the fresh Subject. A later {{cluster-io}}
thread created by one of them inherits the same Subject and can run a token
receiver under it.
The defect is receiver target selection, not token acquisition or client
binding. The proposed fix targets the UGI held by HadoopSecurityContext,
falling back to the login UGI, instead of a pooled thread's current UGI. Here
"UGI" means the Hadoop UserGroupInformation wrapper and "Subject" means the
identity it carries; identity comparisons use Subject identity.
h2. Observed failure
This was observed in a Flink 1.20.5 Kubernetes application-mode job using
ZooKeeper-based high availability and HDFS-backed checkpoint storage. The
primary failure was in the checkpoint storage path:
{noformat}
CheckpointCoordinator.initializeCheckpointLocation
-> FsCheckpointStorageAccess.initializeLocationForCheckpoint
-> HadoopFileSystem.mkdirs
-> DFSClient.mkdirs
{noformat}
After the 24h deadline, the checkpoint-storage client reported InvalidToken for
its startup token even though refresh had obtained a new HDFS token and the
receiver reported success.
Not every Kubernetes application-mode job is affected. The failure is
timing-dependent: a long-lived Hadoop client must capture one Subject/UGI
lifecycle, and a later refresh receiver must run under a different one.
h2. Root cause
The delivery bug is in HadoopDelegationTokenReceiver. It targets
{{UserGroupInformation.getCurrentUser()}} of the refresh thread.
{{getCurrentUser()}} resolves from the thread's effective AccessControlContext
and is not guaranteed to be the UGI held by HadoopSecurityContext or the UGI
captured by a long-lived checkpoint-storage DFSClient. In the observed failure,
the checkpoint client's captured UGI had the login/HadoopSecurityContext
Subject, while the receiver could run under a pooled thread's inherited fresh
Subject.
h3. Acquisition and refresh phases
Token delivery in the JobManager has three phases. Within each phase, the
provider and the JM-local receiver run on the same token-manager thread.
* Bootstrap acquisition: {{ClusterEntrypoint.initializeServices()}} calls
{{DefaultDelegationTokenManager.obtainDelegationTokens()}} on {{main}} before
BlobServer starts. It updates local receivers for early external filesystem
access. The receiver runs after {{freshUGI.doAs(...)}} returns, when current is
login, so this phase is safe.
* ResourceManager startup acquisition: {{ResourceManager.onStart()}} ->
{{DefaultDelegationTokenManager.start()}} -> {{startTokensUpdate()}} runs on
{{default-dispatcher}}. This provider call is where the observed ForkJoinPool
expansion occurs, but the startup receiver itself still runs on
{{default-dispatcher}} after the provider returns, so it does not target the
fresh Subject.
* Scheduled refresh: later updates are submitted through {{cluster-io}}. A
receiver running on a {{cluster-io}} thread that inherited the fresh Subject
resolves {{getCurrentUser()}} to that Subject and can target the wrong UGI.
TaskManager receivers are invoked by {{TaskExecutor}} token updates, not on
{{cluster-io}}; the probe in this issue covers the JM-local receiver path only.
These two startup token acquisitions are separate from the double
{{SecurityUtils.install}} sequence discussed under Scope and boundaries.
h3. How the fresh Subject reaches a pooled thread
The observed Subject-propagation window is the ResourceManager startup
acquisition. The provider runs on a ForkJoinWorkerThread. The fresh Subject
comes from {{KerberosLoginProvider.doLoginAndReturnUGI()}}, which calls
{{UserGroupInformation.loginUserFromKeytabAndReturnUGI(principal, keytab)}}.
That creates a new UGI for the same principal and does not replace the process
login user.
While {{freshUGI.doAs(...)}} is effective, the provider action calls
{{fs.addDelegationTokens(...)}}; the blocking HDFS RPC in
{{CompletableFuture.get()}} enters ForkJoinPool's managed-block path. If no
idle FJP thread is available and the pool has capacity, {{tryCompensate()}}
creates additional dispatcher threads during the blocked RPC. This compensation
does not unblock the RPC; it preserves pool capacity for other work. The
{{Thread}} constructor for each new thread captures the effective
AccessControlContext returned by {{AccessController.getContext()}} into that
thread's {{inheritedAccessControlContext}}; this is a stack-derived snapshot,
not the entire live Java stack, and it includes the Subject active during
{{doAs}}. After {{doAs}} returns, current on the provider thread resolves to
login again, but the newly created threads retain the fresh Subject.
The probe block below combines the observed call chain, Subject identities, and
thread creation. Here HadoopSecurityContext Subject means the Subject of the
UGI held by the installed HadoopSecurityContext; it matches the login Subject
in this probe, but that is not guaranteed in general.
{noformat}
provider thread = default-dispatcher-12
freshUGI Subject = 817147333
login/HadoopSecurityContext Subject = 1333938290
freshUGI.doAs(...)
-> fs.addDelegationTokens(...)
-> HDFS RPC blocked in CompletableFuture.get()
-> ForkJoinPool.managedBlock()
-> tryCompensate()
-> createWorker() [when no idle FJP thread and pool capacity remains]
-> new dispatcher threads inherit the fresh Subject
-> later, one of them triggers creation of a cluster-io thread
current Subject before doAs = 1333938290
current Subject inside doAs = 817147333
current Subject after doAs = 1333938290
FJP-created dispatcher threads:
default-dispatcher-13 / 14 / 15, creator=default-dispatcher-12
later cluster-io thread:
cluster-io-thread-3, creator=default-dispatcher-15
current Subject / JM-local receiver target (getCurrentUser()) = 817147333
login/HadoopSecurityContext Subject = 1333938290
{noformat}
{{fs.addDelegationTokens(...)}} is the provider-side acquisition step: it
obtains the filesystem delegation token or tokens and stores them in a local
{{Credentials}} object. The receiver later delivers that object to the selected
UGI.
{{cluster-io}} is a fixed-size executor; its threads are created lazily on task
submission. In the probe, an FJP-created dispatcher thread later submits a task
to this executor, which creates a cluster-io thread that copies the submitter's
inherited ACC and starts with the fresh Subject. This thread creation is not
the token refresh itself; it only determines the Subject under which the later
scheduled refresh runs. A receiver invocation on that thread resolves
{{getCurrentUser()}} to the fresh Subject and writes credentials there, while
the login and HadoopSecurityContext UGIs remain unchanged and still resolve to
the login Subject shown above.
This is a per-thread inherited security context, not a global mutation of
UserGroupInformation. It is a receiver target-selection mismatch, not a
client-binding error. Clients created under an unrelated transient Subject
require separate rebinding or recreation and are out of scope.
h2. Proposed change
This issue fixes token delivery for the observed long-lived-client case,
primarily checkpoint storage.
* Add an accessor for the UGI held by HadoopSecurityContext.
* In HadoopDelegationTokenReceiver, select a single target UGI: use the UGI
held by HadoopSecurityContext when present; otherwise fall back to
{{UserGroupInformation.getLoginUser()}}. Add refreshed credentials only to that
UGI. Do not target {{UserGroupInformation.getCurrentUser()}}, which may be a
pooled thread's inherited Subject.
* Add unit tests for the receiver's context-first and login-fallback target
selection.
The change makes the delivery target deterministic. It does not rebind a client
already created under a different Subject.
h2. Scope and boundaries
h3. Out of scope
* Preventing HDFS client initialization or blocking RPCs from occurring inside
{{freshUGI.doAs(...)}} or the provider's ForkJoinPool thread.
* Closing or rebinding FileSystem instances created under a transient fresh
Subject; this is a separate client-lifecycle cleanup.
* Delivering credentials to clients created under an unrelated transient
Subject; those clients require separate rebinding or recreation.
* Changing the provider's {{getCurrentUser()}} usage.
{{HadoopFSDelegationTokenProvider}} already acquires tokens inside
{{KerberosLoginProvider.doLoginAndReturnUGI()}}; its remaining
{{getCurrentUser()}} use is a token-required pre-check and renewer-name lookup.
No provider change is required for the observed receiver mismatch; it remains a
separate follow-up if provider-side behavior is observed.
* Double security installation and {{SecurityUtils.install}} idempotency. A
diagnostic build where the second installation was skipped still failed at the
24h deadline. The early/duplicate installation adds lifecycle complexity, but
the probe does not establish it as a cause of the receiver mismatch.
h3. YARN
YARN is not a confirmed reproduction. The YARN samples below were collected
with a Java 8-compatible 1.20.5 probe build; Java 8 remains supported but
deprecated in the 1.20 source line. Across repeated controlled application-mode
probes, some provider windows created no dispatcher thread, while one window
showed a new dispatcher thread during the observation period. However, the FJP
stack probe did not capture a provider-specific
{{tryCompensate}}/{{createWorker}} stack; the observed stacks came from
JobGraph generation and RPC actor paths. The new thread therefore cannot be
attributed to the token provider from the current evidence. One later sample
recorded {{futureDone=false}} immediately before the provider RPC future and
{{true}} after; this shows the future was not yet complete at that point, but
does not prove that {{CompletableFuture.get()}} entered managed blocking,
because it may have completed before the wait. No provider-specific FJP stack
was captured in that sample either. The provider's blocking/compensation state
therefore remains unestablished, as do the per-window FJP state, timing, and
concurrent work as the explanation for the observed difference.
h3. Versions
Before 1.17, the legacy {{HadoopModule}} path added credentials directly to the
login user; the {{HadoopDelegationTokenReceiver}} /
{{UserGroupInformation.getCurrentUser()}} path starts with 1.17. 1.18, 1.19,
and 1.20 share that receiver behavior. 1.18 is not claimed to be immune.
h2. Testing
* Add receiver unit tests for (a) HadoopSecurityContext UGI precedence with
login/current not updated, (b) login fallback when no HadoopSecurityContext is
installed, and (c) the inherited current UGI not targeted.
* Trigger multiple post-startup token refreshes in a long-running job and
verify which executor runs the provider, whether new ForkJoinPool threads are
created, and which Subject the receiver's current user resolves to.
* Repeat the YARN provider-window probe while instrumenting the provider's
{{CompletableFuture}} state; the current YARN sample is summarized under Scope
and boundaries.
* A deployed application-mode job should survive the 24h token deadline and
continue checkpointing without SecretManager$InvalidToken. Verify that the
checkpoint-storage DFSClient receives the new token.
* Run a controlled application-mode probe that verifies the receiver updates
the UGI selected by the new logic rather than a pooled thread's inherited
current UGI; cover the TaskManager receiver path separately.
h2. Acceptance criteria
* A token refresh updates the UGI held by HadoopSecurityContext when present,
otherwise the login UGI, and does not use a pooled thread's inherited current
UGI; a checkpoint-storage client bound to the selected UGI receives the new
token.
* A long-running application-mode job survives the 24h HDFS token deadline.
> Hadoop delegation token refresh can target a different UGI than Hadoop clients
> ------------------------------------------------------------------------------
>
> Key: FLINK-41435
> URL: https://issues.apache.org/jira/browse/FLINK-41435
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Configuration
> Affects Versions: 2.0.0, 1.19.0, 1.20.0, 2.1.0, 2.2.0, 2.3.0, 2.4.0
> Environment: Flink 1.19.0 and later (1.19.x, 1.20.x, 2.0 through
> 2.4), plus master (2.5-SNAPSHOT); Kubernetes and standalone application mode
> with Kerberos and HDFS delegation tokens. The mismatch is more likely when
> FileSystemJobResultStore and the token manager share the cluster-io executor
> (FLINK-27204).
> Reporter: LuChengxiang
> Priority: Major
>
> h2. Summary
> Flink periodically obtains HDFS delegation tokens and distributes them
> through HadoopDelegationTokenReceiver. The receiver adds refreshed
> credentials to UserGroupInformation.getCurrentUser() of the thread running
> the refresh. A long-lived Hadoop client, especially a checkpoint-storage
> DFSClient, may instead use the UGI/Subject captured when that client was
> created. If those identities differ, the token refresh succeeds and the
> receiver reports success, but the new credentials are not delivered to the
> client's UGI. The client continues using its startup token and fails at the
> 24h HDFS token deadline with SecretManager$InvalidToken.
> In the observed Kubernetes application-mode reproduction, the mismatch
> becomes reachable through ForkJoinPool's managed-block compensation. The
> provider blocks in {{CompletableFuture.get()}} on a ForkJoinWorkerThread;
> when no idle FJP thread is available and the pool has capacity, the
> compensation path creates additional dispatcher threads while
> {{freshUGI.doAs(...)}} is effective, and those threads inherit the fresh
> Subject. A later {{cluster-io}} thread created by one of them inherits the
> same Subject and can run a token receiver under it.
> The defect is receiver target selection, not token acquisition or client
> binding. The proposed fix targets the UGI held by HadoopSecurityContext,
> falling back to the login UGI, instead of a pooled thread's current UGI. Here
> "UGI" means the Hadoop UserGroupInformation wrapper and "Subject" means the
> identity it carries; identity comparisons use Subject identity.
> h2. Observed failure
> This was observed in a Flink 1.20.5 Kubernetes application-mode job using
> ZooKeeper-based high availability and HDFS-backed checkpoint storage. The
> primary failure was in the checkpoint storage path:
> {noformat}
> CheckpointCoordinator.initializeCheckpointLocation
> -> FsCheckpointStorageAccess.initializeLocationForCheckpoint
> -> HadoopFileSystem.mkdirs
> -> DFSClient.mkdirs
> {noformat}
> After the 24h deadline, the checkpoint-storage client reported InvalidToken
> for its startup token even though refresh had obtained a new HDFS token and
> the receiver reported success.
> Not every Kubernetes application-mode job is affected. The failure is
> timing-dependent: a long-lived Hadoop client must capture one Subject/UGI
> lifecycle, and a later refresh receiver must run under a different one.
> h2. Root cause
> The delivery bug is in HadoopDelegationTokenReceiver. It targets
> {{UserGroupInformation.getCurrentUser()}} of the refresh thread.
> {{getCurrentUser()}} resolves from the thread's effective
> AccessControlContext and is not guaranteed to be the UGI held by
> HadoopSecurityContext or the UGI captured by a long-lived checkpoint-storage
> DFSClient. In the observed failure, the checkpoint client's captured UGI had
> the login/HadoopSecurityContext Subject, while the receiver could run under a
> pooled thread's inherited fresh Subject.
> h3. Acquisition and refresh phases
> Token delivery in the JobManager has three phases. Within each phase, the
> provider and the JM-local receiver run on the same token-manager thread.
> * Bootstrap acquisition: {{ClusterEntrypoint.initializeServices()}} calls
> {{DefaultDelegationTokenManager.obtainDelegationTokens()}} on {{main}} before
> BlobServer starts. It updates local receivers for early external filesystem
> access. The receiver runs after {{freshUGI.doAs(...)}} returns, when current
> is login, so this phase is safe.
> * ResourceManager startup acquisition: {{ResourceManager.onStart()}} ->
> {{DefaultDelegationTokenManager.start()}} -> {{startTokensUpdate()}} runs on
> {{default-dispatcher}}. This provider call is where the observed ForkJoinPool
> expansion occurs, but the startup receiver itself still runs on
> {{default-dispatcher}} after the provider returns, so it does not target the
> fresh Subject.
> * Scheduled refresh: later updates are submitted through {{cluster-io}}. A
> receiver running on a {{cluster-io}} thread that inherited the fresh Subject
> resolves {{getCurrentUser()}} to that Subject and can target the wrong UGI.
> TaskManager receivers are invoked by {{TaskExecutor}} token updates, not on
> {{cluster-io}}; the probe in this issue covers the JM-local receiver path
> only. These two startup token acquisitions are separate from the double
> {{SecurityUtils.install}} sequence discussed under Scope and boundaries.
> h3. How the fresh Subject reaches a pooled thread
> The observed Subject-propagation window is the ResourceManager startup
> acquisition. The provider runs on a ForkJoinWorkerThread. The fresh Subject
> comes from {{KerberosLoginProvider.doLoginAndReturnUGI()}}, which calls
> {{UserGroupInformation.loginUserFromKeytabAndReturnUGI(principal, keytab)}}.
> That creates a new UGI for the same principal and does not replace the
> process login user.
> While {{freshUGI.doAs(...)}} is effective, the provider action calls
> {{fs.addDelegationTokens(...)}}; the blocking HDFS RPC in
> {{CompletableFuture.get()}} enters ForkJoinPool's managed-block path. If no
> idle FJP thread is available and the pool has capacity, {{tryCompensate()}}
> creates additional dispatcher threads during the blocked RPC. This
> compensation does not unblock the RPC; it preserves pool capacity for other
> work. The {{Thread}} constructor for each new thread captures the effective
> AccessControlContext returned by {{AccessController.getContext()}} into that
> thread's {{inheritedAccessControlContext}}; this is a stack-derived snapshot,
> not the entire live Java stack, and it includes the Subject active during
> {{doAs}}. After {{doAs}} returns, current on the provider thread resolves to
> login again, but the newly created threads retain the fresh Subject.
> The probe block below combines the observed call chain, Subject identities,
> and thread creation. Here HadoopSecurityContext Subject means the Subject of
> the UGI held by the installed HadoopSecurityContext; it matches the login
> Subject in this probe, but that is not guaranteed in general.
> {noformat}
> provider thread = default-dispatcher-12
> freshUGI Subject = 817147333
> login/HadoopSecurityContext Subject = 1333938290
> freshUGI.doAs(...)
> -> fs.addDelegationTokens(...)
> -> HDFS RPC blocked in CompletableFuture.get()
> -> ForkJoinPool.managedBlock()
> -> tryCompensate()
> -> createWorker() [when no idle FJP thread and pool capacity remains]
> -> new dispatcher threads inherit the fresh Subject
> -> later, one of them triggers creation of a cluster-io thread
> current Subject before doAs = 1333938290
> current Subject inside doAs = 817147333
> current Subject after doAs = 1333938290
> FJP-created dispatcher threads:
> default-dispatcher-13 / 14 / 15, creator=default-dispatcher-12
> later cluster-io thread:
> cluster-io-thread-3, creator=default-dispatcher-15
> current Subject / JM-local receiver target (getCurrentUser()) = 817147333
> login/HadoopSecurityContext Subject = 1333938290
> {noformat}
> {{fs.addDelegationTokens(...)}} is the provider-side acquisition step: it
> obtains the filesystem delegation token or tokens and stores them in a local
> {{Credentials}} object. The receiver later delivers that object to the
> selected UGI.
> {{cluster-io}} is a fixed-size executor; its threads are created lazily on
> task submission. In the probe, an FJP-created dispatcher thread later submits
> a task to this executor, which creates a cluster-io thread that copies the
> submitter's inherited ACC and starts with the fresh Subject. This thread
> creation is not the token refresh itself; it only determines the Subject
> under which the later scheduled refresh runs. A receiver invocation on that
> thread resolves {{getCurrentUser()}} to the fresh Subject and writes
> credentials there, while the login and HadoopSecurityContext UGIs remain
> unchanged and still resolve to the login Subject shown above.
> This is a per-thread inherited security context, not a global mutation of
> UserGroupInformation. It is a receiver target-selection mismatch, not a
> client-binding error. Clients created under an unrelated transient Subject
> require separate rebinding or recreation and are out of scope.
> h2. Proposed change
> This issue fixes token delivery for the observed long-lived-client case,
> primarily checkpoint storage.
> * Add an accessor for the UGI held by HadoopSecurityContext.
> * In HadoopDelegationTokenReceiver, select a single target UGI: use the UGI
> held by HadoopSecurityContext when present; otherwise fall back to
> {{UserGroupInformation.getLoginUser()}}. Add refreshed credentials only to
> that UGI. Do not target {{UserGroupInformation.getCurrentUser()}}, which may
> be a pooled thread's inherited Subject.
> * Add unit tests for the receiver's context-first and login-fallback target
> selection.
> The change makes the delivery target deterministic. It does not rebind a
> client already created under a different Subject.
> h2. Scope and boundaries
> h3. Out of scope
> * Preventing HDFS client initialization or blocking RPCs from occurring
> inside {{freshUGI.doAs(...)}} or the provider's ForkJoinPool thread.
> * Closing or rebinding FileSystem instances created under a transient fresh
> Subject; this is a separate client-lifecycle cleanup.
> * Delivering credentials to clients created under an unrelated transient
> Subject; those clients require separate rebinding or recreation.
> * Changing the provider's {{getCurrentUser()}} usage.
> {{HadoopFSDelegationTokenProvider}} already acquires tokens inside
> {{KerberosLoginProvider.doLoginAndReturnUGI()}}; its remaining
> {{getCurrentUser()}} use is a token-required pre-check and renewer-name
> lookup. No provider change is required for the observed receiver mismatch; it
> remains a separate follow-up if provider-side behavior is observed.
> * Double security installation and {{SecurityUtils.install}} idempotency. A
> diagnostic build where the second installation was skipped still failed at
> the 24h deadline. The early/duplicate installation adds lifecycle complexity,
> but the probe does not establish it as a cause of the receiver mismatch.
> h3. YARN
> YARN is not a confirmed reproduction. The YARN samples below were collected
> with a Java 8-compatible 1.20.5 probe build; Java 8 remains supported but
> deprecated in the 1.20 source line. Across repeated controlled
> application-mode probes, some provider windows created no dispatcher thread,
> while one window showed a new dispatcher thread during the observation
> period. However, the FJP stack probe did not capture a provider-specific
> {{tryCompensate}}/{{createWorker}} stack; the observed stacks came from
> JobGraph generation and RPC actor paths. The new thread therefore cannot be
> attributed to the token provider from the current evidence. One later sample
> recorded {{futureDone=false}} immediately before the provider RPC future and
> {{true}} after; this shows the future was not yet complete at that point, but
> does not prove that {{CompletableFuture.get()}} entered managed blocking,
> because it may have completed before the wait. No provider-specific FJP stack
> was captured in that sample either. A controlled YARN probe that instrumented
> {{CompletableFuture.waitingGet}} and {{ForkJoinPool.managedBlock}} produced
> no provider stack, while the same agent produced the expected
> waiting-get/managed-block stack in a self-test. That sample therefore did not
> enter the provider wait path. The new thread observed in another run remains
> unattributed, and the per-window FJP state, timing, and concurrent work
> remain unestablished as the explanation.
> h3. Versions
> Before 1.17, the legacy {{HadoopModule}} path added credentials directly to
> the login user; the {{HadoopDelegationTokenReceiver}} /
> {{UserGroupInformation.getCurrentUser()}} path starts with 1.17. 1.18, 1.19,
> and 1.20 share that receiver behavior. 1.18 is not claimed to be immune.
> h2. Testing
> * Add receiver unit tests for (a) HadoopSecurityContext UGI precedence with
> login/current not updated, (b) login fallback when no HadoopSecurityContext
> is installed, and (c) the inherited current UGI not targeted.
> * Trigger multiple post-startup token refreshes in a long-running job and
> verify which executor runs the provider, whether new ForkJoinPool threads are
> created, and which Subject the receiver's current user resolves to.
> * Repeat the YARN provider-window probe while instrumenting the provider's
> {{CompletableFuture}} state; the current YARN sample is summarized under
> Scope and boundaries.
> * A deployed application-mode job should survive the 24h token deadline and
> continue checkpointing without SecretManager$InvalidToken. Verify that the
> checkpoint-storage DFSClient receives the new token.
> * Run a controlled application-mode probe that verifies the receiver updates
> the UGI selected by the new logic rather than a pooled thread's inherited
> current UGI; cover the TaskManager receiver path separately.
> h2. Acceptance criteria
> * A token refresh updates the UGI held by HadoopSecurityContext when present,
> otherwise the login UGI, and does not use a pooled thread's inherited current
> UGI; a checkpoint-storage client bound to the selected UGI receives the new
> token.
> * A long-running application-mode job survives the 24h HDFS token deadline.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)