Release notes
We recommend always using “tagged” versions of StorageOS rather than “latest”, and to perform upgrades only after reading the release notes.
The latest tagged release is 1.5.4
, available from
Docker Hub as storageos/node:1.5.4
, or via the Helm
Chart
The latest CLI release is 1.2.2
, available from
Github
Upgrading
See upgrades for upgrade strategies.
1.5.4 - Released 2020/08/19
Fixed
- If a CSI volume create request expires before completing, the retry operation may not succeed.
- When detaching NFS volumes, return success if the volume has already been deleted. This avoids Kubernetes retrying the operation indefinitely.
- Don’t overlap /var/lib/kubelet mount. This could result in multiple entries
for
/var/lib/storageos
in/etc/fstab
. - Set fsid_device to false so that NFS does not use local device major/minor numbers for filesystem descriptors.
- Run CSI containers as privileged.
- Pass cluster node selector to NFS pods.
- As an additional safeguard, check that the StorageOS device contains only
zeros in the first 256KB before running
mkfs
.blkid
is also still used to check for existing filesystems. - Reduced writes to etcd for health checks.
- Fixed error “failed to detach volumes” … “resource name may not be empty”.
1.5.3 - Released 2020/01/25
Fixed
- Resolved misrepresentation of internal metrics for licensed clusters.
1.5.2 - Released 2019/12/18
Adds support for Kubernetes 1.17 and CSI 1.2. Also reduces node container size by 70%.
New
storageos/node
container size reduced by removing extraneous contents. This reduces the amount of time required to install and boostrap a new cluster.- Kubernetes 1.17.0 supported, resolving
cannot get resource "leases" in API group "coordination.k8s.io"
permission issue. - Now supports version 1.2.0 of the CSI specification.
Improved
- CSI helpers updated to latest versions.
operator-sdk
dependency updated to version 0.12.- Downgraded some log messages for transient failures from error to info.
Fixed
- Resolved documentation inconsistency between Openshift Marketplace and docs.storageos.com.
- Node label update now retries if the node is not yet known to StorageOS.
1.5.1 - Released 2019/11/26
Fixes a permission error on Kubernetes 1.16 and improves reporting of Pod scheduling failures.
New
- The [cluster-operator] now reports internal metrics suitable for scraping with
Prometheus on
http://0.0.0.0:8686/metrics
. If Prometheus has been installed in the cluster using the [prometheus-operator], thenServiceMonitor
resources will be created to register the endpoint automatically. See [prometheus-setup] for an example Prometheus installation. Currently only the default process metrics are reported but we intend to improve coverage over the next few releases.
Improved
-
Pod scheduling failures are now registered as Pod events to provide contextual error messages. Before, only HTTP status codes could be reported since that was the only information that could be passed via
kube-scheduler
. -
Leader election has been added to [cluster-operator] and “leader-for-life” enabled by default. In the future, this will allow multiple instances of the operator to be running simultaneously. For now, only a single instance is supported. See [leader-election] for more information.
-
The [cluster-operator] now fully implements status phases. This means that Kubernetes and
kubectl
now report the following statuses forStorageOSCluster
resources:Pending
cluster status when the requirements are unmet. New clusters may wait inPending
state while there is an existing cluster running.Creating
cluster status when the provisioning has started.Running
cluster status for a provisioned cluster.Terminating
cluster status when a cluster delete has been initiated.
-
Updated the [cluster-operator] base image to latest
ubi8/ubi-minimal
.
Fixed
-
Fixed permission errors in Kubernetes 1.16:
Failed to list *v1beta1.CSINode: csinodes.storage.k8s.io is forbidden
'events.events.k8s.io is forbidden
-
Deployments using embedded etcd could fail to bootstrap when Kubernetes returns multiple addresses for a node. Previously, the first address in the list was assumed to be the address that the nodes should communicate on. Instead, we now prefer to use the
InternalIP
address, and if that is not found, the first address used as before.
1.5.0 - Released 2019/11/12
Along with a 10x improvement in sequential read performance, version 1.5.0
adds two much-requested features; Pod Locality and Shared Filesytems (RWX).
New
-
Pod Locality improves application performance by placing workloads in close proximity to its data. In the default configuration, and when resources permit, StorageOS will instruct Kubernetes to prefer placing the workload on the node holding the master copy of the data, and then prefer the replicas. Where this is not possible, any other healthy node running StorageOS will be considered.
Pod Locality is enabled by default.
See [Pod locality]({{ ref “docs/concepts/podlocality.md” >}}) for more information.
-
Shared Filesystem support allows StorageOS volumes to be provisioned for read & write access by multiple Pods concurrently. Using the same StorageOS StorageClass, PVC requests can now specify AccessMode
ReadWriteMany
(RWX) for shared filesystem, orReadWriteOnce
(RWO) for a standard volume.Volume features such as replication, encryption and caching can be applied to RWX volumes by using PVC labels and StorageClass parameters in the same way as with RWO volumes.
Each shared filesystem is exposed via a dedicated service, named after the PVC name. The service also exposes Prometheus metrics for the shared volume on
http://<service>:80/metrics
. When Prometheus has been deployed using the Prometheus Operator, ServiceMonitor resources are automatically created for each endpoint.Shared filesystem support is only available when using the CSI driver, which is the recommended deployment method from Kubernetes 1.12 onwards.
Although Shared Filesystem support has been tested by customers in our Beta Programme and we are not aware of any critical issues, we are releasing the feature as a Technology Preview. We encourage all users to test Shared Filesystem support and let us know of any issues or how it can be improved before we recommend it for production workloads.
See Shared filesystems for more information.
Improved
- Read performance for all workload types and block sizes have been improved,
with sequential workloads now 10x faster than version
1.4.0
. - Remote read performance for large block sizes has also been substantially improved with optimizations to reduce round trips, increasing throughput and lowering overall latency. Remote writes for large block sizes have already been optimized.
- Block checksums are now stored with the volume metadata. While not exposed as
features in
1.5.0
, replica consistency can now be quickly proven and corruptions corrected. Existing volumes will not be affected, only newly created volumes. - The dataplane log parser has been replaced and is now more efficient, and handles embedded quotes.
- Telemetry now uses gRPC over TLS. The existing Sentry-based telemetry will be
phased out. Telemetry can be disabled by setting
DISABLE_TELEMETRY=true
.
Fixed
- Encrypted volumes did not always respect the compression setting, resulting in capacity savings being lower than expected. Unlike most storage systems, StorageOS is able to compress data prior to encrypting it.
Other
-
The maximum provisioned capacity of an unregistered cluster (default install) has been reduced from 100GB to 50GB. We have made this change to encourage users to register for the free Developer licence, which grants 500GB of provisioned capacity.
Through cluster registrations we are able to get a better understanding of who is using StorageOS and which events or referrals are most effective. We would love to understand your challenges and any ideas you may have to improve StorageOS. You may unsubscribe from our mailing list at any time.
1.4.0 - Released 2019/09/06
Version 1.4.0
changes the way metadata keys are stored on disk resulting in
improved key locality and thus an increase in IO performance.
IMPORTANT: Due to the nature of this change, the upgrade process includes an
automated conversion process. If upgrading from 1.3.0
or below, please ensure
you follow the upgrade instructions.
If the upgrade process is not followed, version 1.4.0
will refuse to start,
but you may restart using the previous version. Once the upgrade process has
been completed it is not possible to revert to an earlier version.
No new features will be added to the 1.4
series and we will instead release a
1.5.0-rc1
in the next few weeks.
Improved
- Improved performance for sequential IO greater than 4k in size by changing the ordering of metadata keys so that the volume inode is in the most significant bits, followed by the offset.
- Volume deletes are now 5 - 20x faster, and for <10GB volumes should take less than a second.
Fixed
- The Prometheus metrics endpoint no longer displays statistics for volumes that have been deleted.
1.3.0 - Released 2019/07/08
The 1.3.0
release contains further substantial performance improvements,
particularly for workloads that can parallelise IO or when multiple volumes are
performing high rates of IO.
New
- Switched internal device presentation from FUSE to TCMU when the kernel
supports it, and when no conflicting TCMU-based devices already exist. TCMU
improves performance in all use cases, but particularly when queue depths are
greater than 1. TCMU can be disabled by setting
disableTCMU: true
, and it can be forced by settingforceTCMU: true
. WhenforceTCMU: true
has been set and either the kernel does not support TCMU or there are conflicting devices, then StorageOS will refuse to start. - The operator now runs an additional scheduler Pod, which can be disabled by
setting
disableScheduler: true
in the StorageOS CR.
Improved
- The admin user’s username and password are no longer stored in the KV store
and are now always read from the
ADMIN_USERNAME
andADMIN_PASSWORD
environment variables which are populated by the secret fromsecretRefName
. Previously, the environment variables were only read on cluster bootstrap. This caused problems when re-installing a cluster as the previous credentials were cached in the KV store and the credentials set in the environment variables would not work. It is no longer required to clear/var/lib/storageos
between re-installs. - Replication ACK timeouts have be reduced from 15 to 3 seconds to allow retries to be performed sooner. The overall timeout remains the same.
- Telemetry now reports the Kubernetes version and the distribution name if set
in the
k8sDistro
environment variable. This information helps us direct focus on the most relevant platforms. As always, telemetry reporting can be disabled by settingdisableTelemetry: true
. - Low-level dataplane configuration is now included in the diagnostics bundle. The bundle can only be created by the cluster administrator and can be uploaded to assist with support.
- Logging for the Operator has been reworked to allow easier troubleshooting
and consistency with other operators using Kubernetes’
controller-runtime
library.
Fixed
- When creating volumes in stress test conditions, it was possible to be re-allocated the inode of a volume that had not yet finished the deletion process, causing the create to fail.
- When issuing an unmount request using the CSI driver, the operation would fail if the volume was already unmounted. The operation now returns success and allows a remount to succeed.
1.2.1 - Released 2019/05/15
1.2.1 is primarily a bug fix release, but it also includes some performance improvements.
Improved
- Performance tuning on the metadata store, improving performance for sustained writes.
- Metadata checksums now use a pipelined SSE algorithm to improve performance.
- Set SCSI timeouts for StorageOS volumes to 120 seconds.
- More robust handling of devices on first initialisation.
- Volumes created with CSI are now formatted with lazy initialization enabled, reducing the time taken to provision the volume.
- Logging for volume detach operations when fencing a Pod.
- Cluster maintenance logging improved.
Fixed
- Improved handling of in-memory cache during repeated failover events.
- When a volume was being deleted while there was IO pending to an offline remote node, the delete would wait until the IO timed out. IO is now abandoned when a volume is deleted.
- During shutdown, IO to offline remote nodes would be retried, causing shutdown to take longer than needed. Now, unsuccessful IO is marked as failed immediately if a shutdown has been requested.
- The network client is now shutdown before the block presentation layer so that threads with pending IO are immediately unblocked. This helps ensure shutdowns can finish gracefully within the 10 second window set by the orchestrator.
- Node inspect output did not show all configuration information.
- When CSI is enabled, devices are now created on all nodes instead of on demand. This mitigates a timing issue and allows faster failover.
- Replication logs could contain incorrect volume identifiers when multiple syncs are in progress.
1.2.0 - Released 2019/04/17
The 1.2.0
release contains significant performance and resource utilisation
improvements, along with 3 new features:
- Pod fencing
- Volume Encryption
- etcd TLS
New
-
Pod fencing can improve application failover times, particularly for applications deployed as StatefulSets. When the label
storageos.com/fenced=true
is applied to a Pod, that mounts a replicated StorageOS volume, and StorageOS detects that the node running the Pod is offline, it will tell Kubernetes to re-schedule the Pod elsewhere.Without this feature, Kubernetes will by default wait 5 minutes before marking the node offline and performing remedial actions. When applications are deployed as StatefulSets then manual intervention is also required before Pods will be re-scheduled. This behaviour guards against multiple nodes accessing the same volume concurrently as Kubernetes does not have visibility of storage presentations.
Since StorageOS detects and remediates node failure more aggressively (in 30-40 seconds) and it protects against multiple nodes accessing volumes concurrently, the fencing option gives the ability to override the default Kubernetes behaviour.
See fencing for more information.
-
Volume Encryption can now be enabled by setting the
storageos.com/encryption=true
label on a volume. When enabled, data will be encrypted prior to being persisted on disk.See encrypted volumes for more information.
-
Mutual TLS is now supported on external etcd KV backends.
See etcd TLS for more information.
-
When deployed using the StorageOS Cluster Operator, node labels from Kubernetes are visible within StorageOS and available for use in rules. This is particularly useful to mark specific nodes as compute-only. See labels for more information.
Improved
-
Multiple significant performance improvements.
-
Multiple data plane processes were merged into a single
dataplane
binary. This has multiple benefits:- Improved performance by reducing context switches when passing data between processes.
- Easier to instrument and trace.
- Startup and shutdown logic simpler to manage.
-
DirectFS, the network communication layer used by replication and remote volumes, has been largely re-written and now uses a connection pool to reduce memory overhead.
-
Several log formatting and level improvements to provide more relevant and consistent messages.
-
Gossip (serf) port added to the network connectivity heath check.
Fixed
- The CLI command
storageos cluster health
now returns more useful information about the health of the cluster and does not require the cluster ID or join token to be supplied as a parameter. - When changing the cluster log level at run-time using
storageos logs --log-level debug
, the new log level would be applied but an incorrect error would be returned.
1.1.5 - Released 2019/03/29
Version 1.1.5 is an interim bugfix release as we continue to test the upcoming 1.2 feature release internally.
Improved
- Previously we would allocate 5% of free memory to use as cache. On larger systems that were being rebooted, we could use more memory than intended if we start before other applications do. We now set a hard upper-bound of 512MB.
- The cluster ID is now logged more prominently during startup.
- Labels prefixed with
kubernetes.io/
are now accepted. This allows core Kubernetes labels to be inherited and used for scheduling decisions. - A cache flush is now triggered when
fsck
is run on a volume. - Removed noisy log messages from the dataplane stats module.
- Increased log level to
info
for volume create and delete with CSI.
Fixed
- Mounting a volume while the volume was undergoing a series of configuration updates could cause the mount to hang.
- When increasing the number of replicas for a mounted volume, the new replica could get stuck in “provisioned” state. Removing the stuck replica and re-trying would normally clear the issue. The root cause has now been fixed.
- When a volume was being deleted and the data scrub finished, if there was an error removing the volume configuration from the kv store (e.g. kv store unavailable), it would log an error and not retry. Now, the operation is re-tried.
- The soft and hard mode failure detection was incorrectly using the count of failed replicas in the calculation, not the active replicas. When new replicas are being provisioned it’s normal to have more than the requested number of replicas with some of them pending removal. The logic was changed to count healthy replicas instead.
- Fixed excessive log messages when removing a device for a volume that still exists elsewhere in the cluster.
- The UI incorrectly counted the replica stuck in “provisioned” state as available when it did not yet have a consistent copy of the the data.
- When setting the cluster log level via the CLI, an error was returned even though the update suceeded.
- The UI was displaying storage capacity in GB rather than GiB.
- In the UI, the volume delete hover was missing, causing the button to appear disabled.
- Adding a label to a volume in the UI could insert the string “undefined”.
1.1.4 - Released 2019/03/06
Includes a single fix to increase the time we wait for the operating system to create a new block device. On some platforms (cloud providers in particular) the variation can be wider than expected.
Fixed
- Address device creation delays on certain platforms/distributions.
1.1.3 - Released 2019/03/01
Version 1.1.3 is a bugfix and usability release. Please upgrade if you are using external etcd.
New
- There are no new features. Look out for 1.2 coming soon!
Improved
- When using an external etcd KV store and it becomes inaccessible, StorageOS is now more resilient. All volumes remain online and usable but configuration changes are restricted.
- Added a warning in the UI if cluster nodes are running different StorageOS versions. The warning includes a link to the upgrade documentation.
- The cluster ID is now more prominent in the UI. The cluster ID is required for licensing and support.
- UI login failure notifications are now automatically dismissed after a successful login.
- Improved error handling and logging when configuring volumes in the data plane.
- Strict dependency checking when configuring volumes.
Fixed
- When refreshing the volumes page in the UI could return a “volume not found” error.
- When editing volume labels in the UI, there is now a confirmation check before removing a label.
- Fixed a bug that might prevent a volume delete request from being actioned. This would appear as a volume that is stuck in deleting state.
1.1.2 - Released 2019/02/03
1.1.2 improves stability when nodes lose network connectivity intermittenly by introducing a new grace period before performing node recovery operations.
Breaking changes
The environment variables introduced in version 1.1.1
to tune node failure
detection have changed. Previously PROBE_INTERVAL_MS
and PROBE_TIMEOUT_MS
were set to a number of milliseconds. They have now been replaced with
PROBE_INTERVAL
and PROBE_TIMEOUT
, which takes a time duration in string
format, e.g. 500ms
or 2s
. We believe this change gives slightly more
flexibility and is easier to use. Going forward, all duration-based
configuration will use this format.
New
- When network connectivity to a node is confirmed lost, there is now a
configurable grace period before recovery operations begin. The grace period
allows time for the network or node to recover from short interruptions. The
default is set to
30s
. The minimum duration is0s
, and the maximum is90s
. - In the UI, a button has been added to re-send the activation email when requesting a Developer licence.
Improved
-
Logs now use nanosecond precision for timestamps. This helps ensure the correct ordering when logging to an external source such as Elasticsearch.
-
Upgraded Golang compiler to version
1.11.5
. -
Support diagnostics can now be generated while nodes are waiting to join the cluster. This can help troubleshoot dependency or network connectivity issues that are blocking successful startup. This feature is currently only enabled in the API; in a future release it will be exposed in the UI and CLI. To generate the diagnostics bundle, run:
curl -o diagnostics.tar storageos:storageos@localhost:5705/v1/diagnostics/cluster
-
Added more context to error logs when there is a failure to configure the dataplane.
-
Added more logging at debug level when gossip detected a node’s health change.
Fixed
- When CSI is enabled, if a volume was created in a namespace that did not already exist, then the create operation would fail. Now the namespace is created as part of the volume create operation. This matches the behaviour of the Kubernetes native driver.
- When importing Kubernetes ABAC policy files, the jsonl format was incorrectly detected and input was incorrectly parsed as json, which would not validate.
1.1.1 - Released 2019/01/21
1.1.1 contains a critical bugfix in the etcd client library that could cause the StorageOS control plane to become unresponsive when etcd is unavailable.
New
-
Exposed tuning knobs for node failure detection via two new environment variables:
PROBE_INTERVAL_MS
: the interval in milliseconds between node probes. Setting this lower (more frequent) will cause the cluster to detect failed nodes more quickly at the expense of increased bandwidth usage. Defaults to 1000ms.PROBE_TIMEOUT_MS
: the timeout to wait for an ack from a probed node before assuming it is unhealthy. This should be set to 99-percentile of RTT (round-trip time) on your network. Defaults to 3000ms.
The defaults should be appropriate for most environments.
Improvements
- If a node is determined to be offline by memberlist, we now verify by sending a TCP handshake from the scheduler to the node’s replication port. If the handshake is successful we abort the process to mark the node offline.
- If the StorageOS container was killed, we now log at error level instead of debug level.
- The UI now gives user feedback when a diagnostics bundle is being uploaded or downloaded.
- Volume details and
dmesg
output is now collected in the diagnostics bundle. This information helps our engineers diagnose issues more easily. - Internal cleanup of API handlers and middleware. API errors now follow consistent formatting.
- Values for CSI parameters
VolumeId
andVolumePath
are now validated to ensure they are not empty. This is internal to the interaction between the orchestrator and StorageOS and applies to theGetVolumeStats()
CSI call. - Removed spammy “failed to retrieve volume capacity stats” log message when a volume was in the process of being deleted.
- When a node is rebuilt and re-added to the cluster, manual steps must be taken to ensure that a previous (now invalid) config is not re-applied to the new node as it may not have the previous backend data. The log messages now better explain the issue and the steps required to resolve.
- Better error logging when data plane configuration was not applied.
Fixed
- Updated the etcd client library to version 3.3.11, which includes a fix for https://github.com/etcd-io/etcd/issues/9578, a tight loop in the reconnect logic when a watch loses its connection to the etcd server. The loop would cause the control plane to become unresponsive, including not allowing it to respond to healthchecks, which could result in the node being taken offline.
- Fixed an issue in the etcd watcher when we specify an object version to
use as a starting point to watch for change from, but that version no longer
exists in etcd due to scheduled compaction of the etcd datastore. Now, the
error is caught and retried from the latest version. The error was logged as:
etcdserver: mvcc: required revision has been compacted
. - The diagnostics bundle was saving with the extension
.tar.gz
, even though it is not compressed. It now saves as.tar
. - When a node was heavily loaded, it was possible for volume delete requests to get stuck. The delete result processing is now done in a separate thread from the request so that a long request can’t block receiving responses.
1.1.0 - Released 2019/01/04
1.1.0 adds support for CSI 1.0 and includes some minor bug fixes.
New
- CSI 1.0 (available in Kubernetes 1.13 onwards) is now fully supported.
Backwards compatibility with earlier versions of CSI and Kubernetes has been
maintained and is the default. CSI 1.0 support is automatically enabled when
using the StorageOS cluster-operator
on Kubernetes 1.13 and above. If installing manually, set the
CSI_VERSION
environment variable tov1
.
Improvements
- In the UI, the namespaces in the dropdown list are now sorted in alphabetical order.
- Removed spammy log messages for IO size.
- When creating a volume using CSI, the namespace can be set by adding a
volumenamespace
label to the StorageClass.
Fixed
- Cloud provider detection could cause a panic in non-AWS environments that respond on the same metadata ip address that AWS uses. Fixed in aws-sdk-go.
- Refreshing the UI caused the selected namespace to be reset.
1.0.2 - Released 2018/12/10
1.0.2 includes a fix required for Azure AKS and several other fixes and improvements, many to help with running on low-memory or heavily loaded servers.
Improvements
-
Cache allocation has been adjusted for lower memory systems. Allocation sizes are described in architecture. The cache size will never be set to more than 1/3rd of free memory.
-
Increased the maximum number of replicas from 4 to 5.
-
Increased the timeout for the device presentation shutdown to 9 seconds before sending a
SIGKILL
. -
Increased the wait timeout for dataplane startup from 10 to 30 seconds.
-
Added colour-coded health labels in the UI.
-
Rules can now be enabled and disabled in the UI.
-
Improved input validation in the UI.
-
IO error handling log messages are now more informative.
Non-fatal (re-tryable) errors:
Can't resolve replica inode configuration on local director
Can't resolve master inode configuration on the remote director
Can't resolve replica inode configuration on the remote director
The client does not have the appropriate configuration
No connection could be found
Fatal errors:
Generic fatal error (use this if there is no need to be specific)
Can't resolve master inode configuration on the local director
ENOSPC error
Blobs corrupt/missing/not-sane etc
A shm-pipe operation fatally failed
Read only volume
Sentinel value
-
Corrected confusing log message:
poll() returned error on network socket: Success
. We now log:connection closed
. -
Better logging when testing system device capabilities.
-
Better logging for dataplane status errors.
Fixed
- On Azure AKS, StorageOS could fail to restart. This was caused by an error in the system device capabilities verification incorrectly determining that the node didn’t meet prerequisites.
- During startup, it was possible to retrieve an uninitialized value from a dataplane status endpoint, which caused startup to fail. This was only seen on heavily-loaded systems.
- Potential deadlock in dynamic cache resize. This is not known to have occured.
- During a scheduler failover, it was possible to miss a node recovery gossip message and the node would never be marked as recovered.
- It was possible for a volume to get a volume stuck in
syncing
state if a KV watcher was interrupted. Now, full state from the KV store is re-evaluated every 10 seconds. - When creating a volume using CSI, the pool field was ignored.
1.0.1 - Released 2018/11/23
This is primarily a bugfix release to better handle node recovery behaviour when network connectivity is re-established. There are also a few minor bugfixes and general improvements.
Improvements
- Upgraded UI to use Vue CLI 3.
- It is no longer possible for a user to delete their own account, or for an administrator to downgrade their access to a normal user. This ensures that there is always at least one cluster administrator.
- When making volume placement decisions, the scheduler will now apply a lower ranking score to nodes that have come online less than 5 minutes ago. This allows the node to perform most recovery operations prior to accepting new volumes. This will not cause volume placement to be delayed, only that non-recovering nodes will be preferred.
- The cluster diagnostics bundle can now be downloaded via the Web UI in addition to uploading for StorageOS support.
- The volume list in the Web UI now shows more detailed volume health.
- Dataplane metrics now use less memory.
Fixed
- If a node was partitioned as a result of a network failure, it would not always rejoin the cluster when the network was restored. Nodes now attempt to reconnect via gossip every 2 seconds.
- Dataplane process metrics (e.g. memory usage) were not being exposed via
the Prometheus
/metrics
endpoint. storageos logs -f
CLI command was incorrectly reporting all log entries atfatal
level, overwriting the correct log level.- Fixed errors in Web UI when API not yet available.
1.0.0 - Released 2018/11/15
1.0.0 is suitable for production workloads.
New
- Offline nodes can now be removed from management either by using the CLI
node delete
command, or in the UI. This is only available when external etcd is used as the KV store. See decommissioning nodes for more information. - Nodes can now be placed into “Maintenance mode” to disable volume recovery on
the node while it is offline for maintenance. This allows nodes to be
upgraded without triggering potentially unwanted data migrations. Note that
applications accessing StorageOS volumes on the node should also be shutdown.
Alternatively, you may move applications off the node and use
stoageos node drain
to migrate data prior to taking the node offline. - 5% of backend volume capacity is now reserved for indexes and recovery. This helps ensure that there is enough remaining capacity to perform recovery operations such as deleting or moving frontend volumes.
- Prometheus metrics for volume used capacity, and added to the API volume objects and the UI. A placeholder for actual used capacity (after compression) has also been added but is not yet provided.
- In the UI, storage pools and nodes now have an additional details page.
- Added group management to the UI.
- The overprovisioning ratio can now be configured on pools by applying the
label
storageos.com/overcommit
to a positive integer representing the overcommit percentage. By default, overcommit is set to 0. - At startup and every 24 hours, each cluster will make a DNS request to query
the latest StorageOS version and log if a new version is available. The UI
will also notify administrators. This check can be disabled by setting
DISABLE_TELEMETRY=true
. For more information see telemetry. - Since data scrub operations can be intensive, we now limit to 5 concurrent
volume scrubs. This does not affect the number of volumes that may be deleted
in a batch as the scrubs occur asyncronously. Volumes will remain in
deleted
state until the data is scrubbed, then they will not be reported at all.
Improved
- Updated Prometheus metrics names to follow best practices.
- Added process name to dataplane log messages.
- Removed excessive heartbeat messages from debug logging.
- Removed an unneeded mutex from the dataplane debug logging.
- Added protection against a potential hard hang while cleaning up devices by instructing the kernel to abort any IO on the device prior to removing.
- Add
createdAt
andupdatedAt
fields in volume replica API objects. - All API errors now return JSON content type (content remains the same).
- UI improvements to remember state when validation fails.
- UI feedback and help text while requesting a developer licence is more intuitive.
- Better error message when manually applying an invalid licence.
- Volume detail page in the UI has numerous improvements to improve usability, including on smaller screen sizes.
- When the UI was accessed before the node was fully online it would show a
blank page. It now shows the message
Cluster API not yet available, waiting for nodes to join
. - The browser “back” button on the UI now works as expected.
- Label management in the UI now gives context-sensitive examples and improves validation.
- Added network diagnostics and Prometheus metrics to support diagnostics bundle.
- Updated container labels.
Fixed
- When a replica is auto-promoted to master because the application container moved, a cache reset for the volume is now triggered prior to updating the presentation to ensure the cache can never serve stale data.
- Fixed a crash in the replication server during shutdown when a volume scrub operation was taking place on a deleted volume with a non-trivial amount of data.
- In a cluster with hundreds of volumes and nodes continuously rebooting, a race
condition could cause a volume to be “stuck” in syncing state if the sync
started before the dataplane was ready for operation. New replicas are now
created with their health set to
provisioned
, and only enter thesyncing
state once the sync is underway. This allows the operation to be retried if required. - After a restart, a write pointer was starting at the end of the current 64MB chunk. This would waste space if the chunk was not fully used. The pointer is now set to the end of the written data to ensure full utilisation.
- Check-and-set operations in the internal key-value store library were not honouring the TTL requested. This could cause CAS updates to fail when there was above average latency between nodes.
- If a non-admin user tried to cordon/drain a node in the UI it would fail silently. Now the action is disabled for non-admins and the API returns the correct error.
- In the UI, if a previous licensing operation failed, cached data could make it difficult to re-request a licence.
- If a volume was requested but failed due to validation (e.g. not enough
capacity), it would be marked as
failed
and an attempt to re-create would fail with the messageVolume with name 'foo' already exists
. Now the second operation will replace the first and will be re-evaluated. - After startup, pool capacity would only be calculated after the API was ready for use. This caused provisioning requests immediately after startup to fail. Capacity is now calculated prior to the API accepting provisioning requests.
- Slow/broken DNS responses could cause cloud provider detection to delay startup for up to 30 seconds while waiting for a timeout.
- Cluster health CLI and API endpoint no longer report an error when external etcd is used for the KV store.
- When creating new users, usernames are now validated to disallow names in the UUID format used internally.
- In the UI, non-admin users can now view and update their information, including changing their own password.
- Password resets now work correctly and can be used to create a licence for a portal account that was created with a social login.
- If there was an error deleting a rule, the API returned an Internal Server Error rather than a more helpful contextual error.
1.0.0-rc5 - Released 2018/10/01
1.0.0-rc5 is a major update, with multiple bug fixes, performance and usability improvements as we get closer to removing the RC label.
Breaking changes
- Between rc4 and rc5 the location of the lock that governs configuration changes has moved. For maximum safety, shutdown StorageOS on all nodes prior to upgrading to rc5. If this is not feasible, contact [email protected] for further instructions.
- Prometheus metric names and types have changed to adhere to best practices. If you have existing dashboards they will need to be updated.
New
- User, group and policy management added to the Web UI.
- A custom or self-hosted cluster discovery service can now be specified by
setting the
DISCOVERY_HOST
environment variable to the hostname or ip of the running service. The source code for the discovery service is open and available on GitHub. Note that the cluster discovery service is optional and supplied as a convenience. - The
KV_ADDR
environment variable now supports specifying multiple hostnames or addresses of external etcd servers. Hostnames and addresses should be comma-separated and can include the port number (required if not2379
, the etcd default). - Passes CSI 0.3 (latest) conformance tests.
- RHEL/Centos has a limit of 256 devices per HBA. The API now returns
cannot create new volume, active volumes at maximum
when this limit has been reached. - New network connectivity diagnostics to help diagnose potential firewall
issues. With the CLI (
storageos cluster connectivity
) or API (GET /v1/diagnostics/network
), all required connectivity will be verified. - Without a licence, StorageOS has all features enabled but provisioned capacity is now limited to 100GB. Once registered (for free, via the Web UI), capacity increases to 500GB.
Improved
-
Increased overall performance by reducing context switches in the IO path, resulting in lower latency for all IO.
-
Increased replication performance by optimising the parallel writes to multiple destinations. With >1 replicas this will roughly double replication performance, but it also reduces the overhead of adding additional replicas to a volume.
-
Time taken to detect and recover from a node failure reduced from 45-70 seconds to <15 seconds.
-
Improved the internal distributed lock mechanism to ensure correct and deterministic behaviour in the majority of failure scenarios.
-
Implemented basic check-and-set (CAS) on control plane operations where consistency is required.
-
Stop all control plane state evaluations if KV store is unavailable for more than the distributed lock TTL (5 seconds).
-
Ensure node does not participate in state evaluations when it has been set as unschedulable with
storageos node cordon
orstorageos node drain
. -
Internal library change from Serf to Memberlist. This helps simplify node failure detection.
-
Control plane state evaluations are now performed serially with an interval of once per second. This reduces the load on the cluster during bulk or recovery operations. Previously, multiple workers may try to apply the same changes, causing unnessessary validation to take place. A Prometheus histogram has been added to track duration of each state evaluation. Under normal operation it takes 10-50ms.
-
Node capacity now includes capacity from all devices made available for use by StorageOS.
-
Volume health management and reporting is improved. Health is now defined as:
healthy
: Only if the volume master and requested number of replicas are healthy.syncing
: If any replicas are in the process of re-syncing.suspect
: The master is healthy, but at least one replica is suspect or degraded.degraded
: If the master is suspect/degraded or the master is healthy but at least one replica is dead.offline
: If the volume is offline, typically because the master is not available and there were no healthy replicas to failover to.decommissioned
: If a volume is being deleted the health will be marked as decommissioned.
-
Prometheus metrics have been overhauled to provide more friendly and useful statistics.
-
Internal library change from vue-resource (deprecation notice) to Axios for handling ajax requests in the Web UI.
-
Warnings are now logged when volumes can not be provisioned due to licensing constraints.
-
Internal logging improvements to ensure that runtime changes to log levels or filters maintain consistency across multiple consumers.
-
Diagnostic requests now support JSON requests via the “Accepts” header, defaulting to the current tar archive response. Internally, collection from multiple nodes is now streamed with a timeout of 10 seconds to prevent an unresponsive node from blocking the response.
-
Improved description and flow of cluster diagnostics upload.
-
Web UI volume detail page re-designed.
-
Web UI theme tweaks: list view, pagination, headers, in-place edit.
-
Upgrade version of internal messaging library (nats) and restructure implementation.
-
Ensure the API supports label selectors on all object listings and improve internal code consistency and tests.
-
Ensure control plane workers and gRPC connection pool are shutdown cleanly prior to shutting down the data plane to ensure operations have completed and to reduce warnings in logs.
-
Explicitly shutdown the data plane gRPC handlers in the replication client prior to shutting down the rest of the data plane. This protects against a potential issue that could lead to unclean shutdown.
-
The lun ID is now computed dynamically. This allows support for devices across multiple HBAs in the future, which will remove the limitation of 256 volumes on RHEL7/Centos7 clusters and other systems with kernel 3.x.
-
SCSI lun support now supports co-existence with other block storage providers
-
Bulk volume creation could cause excessive validation warning messages in the logs do to concurrent configuration requests to the data plane. Now, if the volume already exists or a CAS update fails, a conflict error (rather than generic error) is returned to the control plane. This allows the control plane to only log a warning or error message if the intended action failed.
-
Ongoing log message readability and tuning of log levels.
Fixed
- Filesystems now behave correctly when the underlying data store is out of capacity. Previously the filesystem would appear to hang as it would retry the operation indefinitely. Now, the filesystem will receive a fatal error and will typically become read-only.
- Listing volumes bypassed policy evaluation so a user in one namespace was able to view volume details in another namespace. Create, update and delete were not affected and required correct authorization.
- Deletion of encrypted volumes did not work correctly.
- Product version details were not always displayed correctly.
- Licence upload could fail with error:
licence and actual array UUIDs do not match
. - Erroneously reported an error when removing data for a volume that was already deleted.
- Web UI icons no longer load from the Internet and now display in disconnected environments.
- Web UI capacity now uses gigabytes to be consistent with CLI.
- API error when running
storageos cluster health
was suggesting to usestorageos cluster health
to diagnose the errror. - API for retrieving the Cluster ID is no longer restricted to administrators only, but any authenticated user.
- Rules against namespaces with common prefixes no longer clash.
- It was possible to create more than one namespace with the same name.
- Fixed
fork/exec /usr/sbin/symmetra: text file busy
, caused by the operating system not having closed the file for writing before it is executed. - Fixed
docker volume rm <volname>
returning 0 but not deleting the volume correctly. - If a volume has a single replica and the master fails when the replica is in
syncing
state, the volume is now marked asoffline
until the master recovers. Previously, the volume remained insyncing
state indefinitely. - If a volume with no replicas goes offline and then recovers, the volume was
not marked as
healthy
. - The CLI was reporting replicas that were not active, creating inaccurate volume counts.
- On a new volume with no replicas with the master on node1, if the volume was mounted on a node that does not hold the volume master (node2) and then node1 drained, then the new volume created on another node will not be mountable.
- Prometheus
/metrics
endpoint always returns text format with correct text encoding, even if binary output was requested. This matches the Prometheus 2.0 policy to deprecate the binary format, and allows the endpoint to work with other collectors that support the text format, e.g. Telegraf. - Fixed an unclean shutdown issue in the data plane volume presentation where configuration of a volume could still be attempted while a shutdown was in progress. A lock is now taken to ensure order and to refuse configuration while a shutdown is in progress. This only occured during shutdown and did not affect data consistency.
- Fixed an unclean shutdown issue in the replication client where references to the client configuration could be cleared while IO was still in progress. A reference counter has been implemented to ensure that all operations have completed prior to the server entry being removed. This only occured during shutdown and did not affect data consistency.
- Ensure all internal IO acknowledgements are received or a timeout reached prior to shutting down the replication server. This fixes an issue where a shutdown or fatal error could cause the server to crash as it was being shutdown cleanly. This only occured during shutdown and did not affect data consistency.
- Fixed an issue in the replication client connection handler that in some situations could allow the client to re-use a connection while it was still being prepared for re-use. This could lead to a replication error or volumes stuck in “syncing” state.
- No longer logs an error when deleting a volume with no data.
1.0.0-rc4 - Released 2018/07/18
Multiple bug fixes and improvements, and improved shutdown handling.
New
- Available capacity is now used in volume placement decisions. This should allow for a more even distribution of capacity across the cluster, especially in clusters with large volumes.
- Licenses can now be applied from the Web UI.
Improved
- The shutdown process has been improved within the data plane, speeding up the removal of devices and reducing the risk that the container runtime or orchestrator will forcefully stop StorageOS.
- IO operations on volumes being deleted now return a fatal error so that the operation is not retried and can fail immediately.
- During node filtering for volume scheduling decisions, it was possible to return duplicate nodes as candidates for the volume placement. This resulted in slightly unbalanced placement across the cluster.
- Volume replica removal didn’t prioritise syncing volumes over healthy. This meant that a fully-synced replica might be removed instead of one that is not yet synced, in cases where the number of replicas was increased and then immediately reduced.
- Log verbosity has been reduced at info level. Set
LOG_LEVEL=debug
for more verbose logging.
Fixed
- UI labels overlapped when window resized in Firefox.
- You can now provision the licenced amount, rather than up to the licenced amount fixing: “cannot provision volume with size 999 GB, currently provisioned: 1 GB, licenced: 1000 GB”.
1.0.0-rc3 - Released 2018/07/02
Multiple improvements based on customer feedback.
New
- StorageOS can now be installed alongside other storage products that make use of the Linux SCSI Target driver.
- The CLI now checks for version compatibility with the API.
Improved
- Volume deletion now uses the scheduler’s desired state processing rather than the previous imperative operation. This fixes an issue where deletes could be stuck in pending state if the scheduler loses state (e.g. from a restart) while the operation is in progress. Now the operation is idempotent and will be retried until successful.
- Volume placement now distributes volumes across nodes more evenly.
- CSI version 0.3 (latest) is now fully supported. Additionally, improvements to CSI include how the default filesystem is determined, read-only mounts, and better checking for volume capabilities.
- Internal communication now times out after 5 seconds instead of 60. This allows retry or recovery steps to initiate much quicker than before. This timeout only affects inter-process communication on the same host, not over the network to remote hosts.
- Added a “degraded” state to the internal health monitoring. This allows a recovery period before marking a node offline, which then triggers a restart. This improves stability when the KV store (internal or external) is undergoing a leadership change.
- Minor improvements to the UI notifications and error messages.
- Upgraded controlplane compiler to Golang 10.0.3 (from 1.9.1).
Fixed
- Online device resizing now works with SCSI devices. Note that
resize2fs
still needs to be run manually on the filesystem, and we are working on making this step automated. - When deleting data from a volume, some metadata was not always being removed. This meant that volumes with frequently changing data could use more capacity than allocated.
1.0.0-rc2 - Released 2018/05/31
Single fix to address provisioning issue in Amazon AWS.
Fixed
- Increased time to wait for device to appear, causing volume creation to fail. This was encountered on Xen-based RHEL/Centos VMs running in Amazon AWS.
1.0.0-rc1 - Released 2018/05/25
The 1.0 release series is focussed on supporting enterprise workloads, with numerous new features and improvements to performance, stability and maintainability.
Breaking changes
- Several internal data structures were changed and are not compatible with previous versions. For this reason upgrades from version 0.10.x and earlier are not supported.
- The API endpoint
/v1/metrics
has been replaced with/metrics
to conform with Prometheus best practices.
New
- Volume presentation has changed to use the SCSI subsystem via the Linux SCSI Target driver. Previous versions used NBD where available, or FUSE where it wasn’t. Using the new volume presentation improves performance on the RHEL platform where NBD is not available. This feature is available for all major distributions and is widely used. For more information, see system configuration
- Internally, the StorageOS scheduler has switched to using level-based state handling and the gRPC protocol. This allows the scheduler to make assertions about the current state, rather than relying on events that can be missed. The scheduler now behaves in the same way as Kubernetes controllers do; by evaluating the current state, calculating adjustments, sharing desired state and allowing individual components to apply differences. For more information on level triggered logic, see Edge vs Level triggerd logic
- Encyption at-rest is available on a per-volume level by setting the label
storageos.com/encryption=true
. In Kubernetes, volume encryption keys are stored in Secrets. Elsewhere, keys are stored in etcd, the key-value store used internally within StorageOS. No other configuration is required. - Container Storage Interface (CSI) version 0.2 compatibility. CSI is a specification for storage providers that enables StorageOS to support any orchestrator that supports CSI. Currently this includes Kubernetes and derivatives, Mesos and Docker. Over time CSI will replace the native driver in Kubernetes, though the existing v1 API will remain.
- Multiple devices on a node can now be used, and StorageOS will shard data across them.
- Node maintenance command such as
storageos node cordon
andstorageos node drain
aid upgrades by live-migrating active volumes prior to node maintenance. - Node labels can now be used for volume anti-affinity, allowing StorageOS to make sure a master volume and its replicas are in separate failure domains.
- Cloud provider failure domains are now auto-set as labels for nodes deployed in Azure, AWS or GCE.
- Administrators can download or upload to Storageos, support, cluster information and log files. This is currently only available via the Web interfce or the API.
- Kubernetes mount option support (requires Kubernetes 1.10).
Improved
- Performance has improved significantly throughout and has been thoroughly tested on a variety of workloads. Specific areas include optimisations for larger block sizes and improvements to the caching engine.
- Volumes will go read-only shortly before the underlying device runs out of space. This allows the filesystem to handle errors gracefully and can protect against corruption.
- Pools have been overhauled to be dynamic, making use of label selectors.
- Increased number of volumes per node.
- Invalid labels now generate a validation error.
storageos.com/
namespaced labels must validate against known labels.
Fixed
- Mapping of compression and throttle labels.
- Remove hard connection limit on replication services.
docker ps
shows unhealthy when everything is fine.- Throttle impact was too small to be noticed.
- Volume size ‘0’ produced 10GB volume.
- Re-adding user to a group creates a duplicate entry.
- Node and pool capacity stats were sometimes wrong.
Previous Releases
Upgrades
Due to breaking changes between 0.9.x and 0.10.x upgrading is not possible. Instead we recommend that you provision a new cluster and migrate data.
Please update to the latest CLI when installing 0.10.x.
0.8.x -> 0.9.x
Start-up scripts should be updated to use the new cluster discovery syntax
Do not mix a cluster with 0.8.x and 0.9.x versions as port numbers have changed. This may cause cluster instability while nodes are being upgraded.
0.7.x -> 0.8.x
Due to the nature the KV Store change there is no upgrade method from 0.7.x to
0.8.x+. Our recommendation is to create a new cluster, paying attention to the
new parameters (CLUSTER_ID
and INITIAL_CLUSTER
). Note that CLUSTER_ID
and
INITIAL_CLUSTER
have been replaced by JOIN
in 0.9.x onwards.
0.10.0 - Released 2018/02/28
The 0.10.0 version focusses on stability and usability as we get closer to GA, but also adds a number of new features (UI, Prometheus metrics, log streaming).
Breaking changes
We took the decision to bundle as many known breaking changes into this release in order to reduce pain later.
-
The API endpoint
/v1/controllers
has been replaced with/v1/nodes
. This removes the concept of storage controllers being different objects from storage clients. Now both are simply nodes, with thestorageos.com/deployment
label used to configure the deployment type. Deployment types can becomputeonly
andmixed
are supported, withmixed
being the default.Older versions of the CLI should be mostly backwards compatible but some features such as node and pool capacity will not work.
-
The label format for configuring StorageOS features has changed from
storageos.feature.xyz
tostorageos.com/xyz
. This format is more familiar to Kubernetes users. Using the old format will still work until 0.11 when conversion will be removed. In 0.10 a deprecation notice will be issued. -
The environment variable used to add labels to nodes at startup time has changed from
LABEL_ADD
toLABELS
. Multiple labels can be comma-separated:LABELS=storageos.com/deployment=computeonly,region=us-west-1
New
-
There is now a Web interface, available on all cluster nodes on the API port (
5705
). To access, point your web browser to http://ADVERTISE_IP:5705. where ADVERTISE_IP is the public ip address of a StorageOS node. Unless you have changed the default credentials, login with userstorageos
, passwordstorageos
. See: https://docs.storageos.com/v1.x/docs/reference/gui/ -
Prometheus stats are exported on each cluster node at http://ADVERTISE_IP:5705/v1/metrics. Prometheus 2.x is required.
-
The CLI can now stream realtime logs from the active StorageOS cluster with
storageos logs -f
. Logs are aggregated from all cluster nodes over a single connection to the API. Future releases will add support for filtering and controlling the log level. See: https://docs.storageos.com/v1.x/docs/reference/cli/logs/ -
The location of the StorageOS volumes can now be configured on a per-node basis by setting the
DEVICE_DIR
environment variable on the StorageOS node container.This is especially useful when running Kubernetes in a container and you are unable or unwilling to share the default
/var/lib/storageos/volumes
location into thekubelet
container. Instead, share thekubelet
plugin directory into the StorageOS container, and setDEVICE_DIR
to use it. Hovever, Kubernetes 1.10 will be required in order forkubelet
to use a directory other than/var/lib/storageos/volumes
.For example:
JOIN=$(storageos cluster create) docker run --rm --name storageos \ -e HOSTNAME \ -e DEVICE_DIR=/var/lib/kubelet/plugins/kubernetes.io~storageos/devices \ -e JOIN=${JOIN} \ --pid host \ --privileged --cap-add SYS_ADMIN \ -v /var/lib/kubelet/plugins/kubernetes.io~storageos:/var/lib/kubelet/plugins/kubernetes.io~storageos:rshared \ --device /dev/fuse \ --net host \ storageos server
Improved
soft
volume failure mode will now tolerate the replica being offline (for example during a node reboot), if there is only one replica configured. To ensure there are always two copies of the data, usehard
mode with a single replica, or use two replicas withsoft
mode.- Ensure volumes can only be mounted with the correct underlying filesystem.
You will now get a validation error if you try to mount an ext3 formatted
filesystem as ext4.
ext4
is the default if not specified when mounting an unformatted volume.
Fixed
- Ensure cache is invalidated after mkfs. Fixes mount error 32 that could occur with Kubernetes or OpenShift on Centos/RHEL.
- Memory leak in replication client led to excessive use over time.
- Standardised all paths to 256 byte length to support user-configurable
DEVICE_DIR
. - Fixed clean shutdown issue where filesystem could start new threads while shutting down. This was only observed in stress tests.
- Shutdown signal handling improved, fixing
Transport endpoint is not connected
on Centos / RHEL in certain restart situations.
0.9.2
The 0.9.2 fixes a memory leak, so upgrading is recommended. It also reduces the time taken to delete volumes.
More work has gone into logging, with API endpoints to control verbosity and
log filtering of running nodes individually or cluster-wide. Version 0.9.4 of
the CLI provides the storageos logs
subcommand to control settings remotely.
New
- API endpoints for logging configuration
Improved
- Remove cgo dependencies from controlplane
- Better error message when the KV store is unavailable
- Reduced memory footprint of controlplane by 67%.
Fixed
- Memory leak introduced in 0.9.0 fixed
0.9.1
The 0.9.1 release includes a fix for the mount deadline exceeded
error, where
a volume mount fails after a long delay.
The logging susbsystem has also been overhauled to reduce noise and to allow filtering messages by log level and component. We expect to build on this in upcoming releases to further improve diagnostics.
New
-
Failure modes can be specified to control behaviour in specific failure scenarios. Full documentation to be added to the docs site soon.
hard
will enforce the desired number of replicas and take a volume offline if the replica count is not met. In practice, the volume should only go offline if there isn’t a suitable node to place a new replica on. For example, if you have configured 2 replicas in a 3 node cluster and a node goes offline.alwayson
will optimise for availability and will keep a volume online even if all replicas have failed.soft
is set by default, and will only take a volume offline if the number of replicas falls below the Failure tolerance (default: number of replicas - 1). This will allow a volume with 2 replicas in a 3 node cluster to remain active while a node reboots.
You can select the failure mode through the labels:
storageos volume create --label storageos.com/replicas=2 --label storageos.com/failure.mode=alwayson volume-name
-
Log filtering allows finer-grained control of log messages so that debug-level messages can be enabled only on some components. Filtering is controlled by setting
LOG_FILTER=cat1=level1,catN=LevelN
. For example, to enable debugging on only the etcd category:- Set
LOG_LEVEL=debug
to enable debug logs. This must be set to the lowest level requested in the filter. - Set
LOG_FILTER=cp=info,dp=info,etcd=debug
to set controlplane and dataplane logs back toinfo
level, thenetcd
todebug
.
- Set
-
Enabled profiling of the controlplane when DEBUG=true is set. The endpoint is availble at http://localhost:5705/debug/pprof.
-
Debug endpoints to monitor and control KV store leadership (DEBUG=true must be set).
- HTTP GET on http://localhost:5705/debug/leader returns true/false if the node is the cluster leader.
- HTTP PUT on http://localhost:5705/debug/leader/resign will cause leader to resign, triggering a new election.
- HTTP PUT on http://localhost:5705/debug/leader/run will cause node to run for cluster leadership.
These actions are used primarily for automated testing where we introduce instability into the cluster to ensure there is no service disruption.
Improved
- More instructive log messages if unable to create filesystem.
- Improve check for whether a device is using NBD for presentation.
- Log whenever a dataplane process restarts.
- Updated etcd libraries to v3.3.0-rc.0.
- Etcd tuning (TickMs = 200, ElectionMs = 5000)
- Updated other dependencies to latest stable releases.
- Dataplane, etcd, gRPC and NATS now use same log format as the controlplane.
- More user-friendly error message when
ADVERTISE_IP
is invalid.
Fixed
- Fixed a bug that caused mounts to fail with
Failed to mount: exit status 32
. This was caused by the volume being marked as ready before the device was fully initialised, and the mount starting before the initialization completed. - Volume names are no longer lowercased and keep the requested case. This fixes an issue with Docker EE with mixed-case volume names.
- Volume delete on non-existent volume now returns HTTP 404 instead of 500.
- Do not reset node health if internal healthcheck returns invalid response. Instead, retry and wait for valid response message.
0.9.0 - Released 2017/11/17
This release focusses on usability and backend improvements. It builds on the embedded KV store from 0.8.x and improves the bootstrap process.
Enhanced stress testing and performance benchmarking have also led to a number of improvements and better failure handling and recovery.
Deprecation notices
While in Beta there may be changes that break backwards compatibility. GA releases will strive to preserve compatibility between versions.
- The
INITIAL_CLUSTER
andCLUSTER_ID
environment variables have been replaced withJOIN
and they should no longer be used to bootstrap a new cluster.JOIN
provides a more flexible mechanism for creating new clusters and expanding existing clusters. Errors or deprecation notices in the log at startup will help verify that the correct environment variables are used. - The API object format for users has changed to make it consistent with other objects. Version 0.9.0+ of the StorageOS CLI must be used to manage users on a version 0.9.x cluster.
New
- Cluster can now start as soon as the first node starts, and any number of
nodes can join the cluster. The
JOIN
environment variable replacesCLUSTER_ID
andINITIAL_CLUSTER
and is more flexible, allowing administrators to combine methods for bootstrapping and discovering clusters. See https://docs.storageos.com/docs/install/reference/clusterdiscovery - Online replica sync for new and replacement replicas. This allows for writes to complete on the required number of replicas immediately instead of waiting for a new replica or a replacement replica to be fully synced.
- Network ports now use a contiguous range (5701-5711/tcp, 5711/udp)
- State-tracker added to frontend filesystem. This allows greater flexibilty for backend failure recovery.
- Prometheus endpoint for exposing internal metrics. This version includes API usage metrics, but will be extended in an upcoming release to include detailed volume, pool and node metrics.
- Reports are sent to Sentry if processes within the StorageOS container are
stopped unexpectantly. These reports help us improve the stability of the
product, but they can be disabled by setting the setting the
DISABLE_TELEMETRY
environment variable to true.
Improved
- Enforced ordering for some types of read/write operations to ensure deterministic behaviour
- Ensured all components shutdown cleanly in normal situations and recover properly when clean shutdown not done.
- Added gRPC to dataplane for synchronous internal messaging
- User API endpoint now has a format consistent with other endpoints.
/users
now returns a list of users, and a POST/PUT to/users/{id}
accepts a user object. - Document minimum server requirements
- Document node selector for volume create operations
- Document quality of service (QoS) and protection from noisy neighbors
- Document volume placement hints
- Separate usage metric endpoints for internal (Dev/QA) and release (Public)
- Health endpoints show
changedAt
andupdatedAt
to help detect flapping services - Unmount volumes through CLI and API
- Removed false-postive errors and warning from logs
- Improve signal handling
LOG_LEVEL
passed to dataplane components
Fixed
- Clear mount lock after ungraceful node shutdown
- etcd error:
mvcc: database space exceeded
- NBD cleanup during ungraceful node shutdown
- No longer possible to have two default pools
storageos cluster health
did not return useful information while the cluster was bootstrapping- Node master and replica statistics were not getting updated after a node failure
Known issues
- It is currently not possible for a node to leave the cluster completely. If
the StorageOS container is stopped and/or removed from a node then the node
will be detected as failed and it will be marked offline, but there is no way
to remove the node from the list.
storageos node rm
will be added before GA along withstorageos node cordon
to disable scheduling new volumes on the node, andstorageos node drain
to move volumes to other nodes prior to maintenance or removing the node from the cluster. storageos volume mount <vol> <mountpoint>
does not work on Managed Plugin installs. Volumes mount into containers correctly using Docker.- Docker can only access volumes created in the
default
namespace. - Clients mounting volumes from RHEL7/CentOS 7 will experience degraded performance due to the absence of the NBD kernel module on those platforms.
0.8.1
Improved
- NBD device numbers now start at 1 instead of 0 to defend against default values
0.8.0 - Released 2017/08/29
This is our first feature release since launching our public beta, and it focusses on feedback from users. As always, please let us know how you are using StorageOS, what problems it is solving for you and how it can improve. Join our Slack channel!
New
- Embedded KV Store based on etcd to further simplify deployment and ongoing cluster management. Support for external Consul KV clusters has been deprecated, though external etcd clusters are now possible, but not yet documented and supported.
- Cluster discovery service to help bootstrap clusters. Allocate a new cluster
with
storageos cluster create
and pass the cluster identifier to each StorageOS node in theCLUSTER_ID
environment variable to allow nodes to discover the cluster without specifying hostnames or ip addresses. - Access control policies can restrict access to volumes and rules created within a namespace.
- User and group management allows multiple users to be created and then used to apply access policy by group membership or named account.
- Anonymized usage metrics are collected and sent to StorageOS to help us better
understand usage patterns so we can focus our efforts accordingly. Metrics
can be disabled by setting the
DISABLE_TELEMETRY
environment variable totrue
. - Location-based scheduling allows administrators to specify scheduling
constraints on volumes at creation time. This provides a simple mechanism to
influence data placement. (e.g. The volume’s data may only be stored on nodes
which have their environment label set to
production
) - Cluster health reporting with CLI (
storageos cluster health
)
Improved
- Graceful behaviour when communication blocked by firewalls
- Docker integration now supports ext2/3/4, btrfs and xfs
- Environment variable validation
- Selectors for rules
- Use default namespace when not specified
- Internal volume performance counters
- API can report health while waiting for cluster to initialize
Fixed
- Better behaviour when communication blocked by firewalls
- Ensure namespaces are unique when creating
- Volume create should give error when size=0
- CLI/API filters not working as expected
- Can’t edit a pool with the CLI
- Excessive logging on network timeouts
Known issues
- Once a 0.8.0 cluster has been established, it is currently not possible to add or remove members. We expect this functionality to come in 0.8.1, and welcome feedback on how this should behave.
storageos volume mount <vol> <mountpoint>
does not work on Managed Plugin installs. Volumes mount into containers correctly using Docker.- Docker can only access volumes created in the
default
namespace. - Clients mounting volumes from RHEL7/CentOS 7 will experience degraded performance due to the absence of the NBD kernel module on those platforms.
0.7.10
This release continues our focus on stability and better test coverage across a growing suite of platforms and usage scenarios.
New
- Keepalives for replication though firewalls
Improved
- Graceful shutdown internal cleanup
- User authentication internals made extensible to other auth providers
- Internal changes to allow different log levels per module
Fixed
- Removed log entries where events are raised and logged again
- Consolidated log messages when KV store is not available
0.7.9
New
- Maintenance mode to allow partial startup when the KV store is not ready
Improved
- Better write parallelization performance
Fixed
- Issue with volume deletion
0.7.8
New
- Node statistics reporting
- Compression (in-transit and at-rest) enabled by default
Improved
- Refactored startup process
- Better volume distribution across cluster
- Relaxed volume naming requirements to support Docker dynamic volumes
- More instrumentation counters for volume stats
- Improved data plane error handling and retry logic
- Replication enhancements
- Ongoing log-level tuning to reduce noise
Fixed
- Issue removing metadata on volume destroy
- Improved compression performance
- Default volume size now consistent across CLI/API/Docker/Kubernetes
- KV Store documentation (thanks @wcgcoder)
0.7.7
- Fix issue with replication after node failure
0.7.6
- Log formatting improvements
- Do not mark a recovered controller available immediately, wait to detect flapping
- Mount lock improvements
- Docker should use ext4 as default filesystem
0.7.5
- Remove bash
0.7.4
- Log formatting improvements
- Add support for ext2, ext3, xfs and btrfs filesystem types
- Don’t allow pool delete with active volumes
0.7.3 and earlier
- Unreleased private betas