This is the second release candidate for Luminous, the next long term stable
release. Please note that this is still a *release candidate* and not
the final release, and hence not yet recommended on production clusters,
testing is welcome & we would love feedback and bug reports.
Ceph Luminous (v12.2.0) will be the foundation for the next long-term
stable release series. There have been major changes since Kraken
(v11.2.z) and Jewel (v10.2.z), and the upgrade process is non-trivial.
Please read these release notes carefully.
Major Changes from Kraken
-------------------------
- *General*:
* Ceph now has a simple, built-in web-based dashboard for monitoring
cluster status.
- *RADOS*:
* *BlueStore*:
- The new *BlueStore* backend for *ceph-osd* is now stable and the new
default for newly created OSDs. BlueStore manages data stored by each OSD
by directly managing the physical HDDs or SSDs without the use of an
intervening file system like XFS. This provides greater performance
and features.
- BlueStore supports *full data and metadata checksums* of all
data stored by Ceph.
- BlueStore supports inline compression using zlib, snappy, or LZ4. (Ceph
also supports zstd for RGW compression but zstd is not recommended for
BlueStore for performance reasons.)
* *Erasure coded* pools now have full support for *overwrites*,
allowing them to be used with RBD and CephFS.
* The configuration option "osd pool erasure code stripe width" has
been replaced by "osd pool erasure code stripe unit", and given the
ability to be overridden by the erasure code profile setting
"stripe_unit". For more details see "Erasure Code Profiles" in the
documentation.
* rbd and cephfs can use erasure coding with bluestore. This may be
enabled by setting 'allow_ec_overwrites' to 'true' for a pool. Since
this relies on bluestore's checksumming to do deep scrubbing,
enabling this on a pool stored on filestore is not allowed.
* The 'rados df' JSON output now prints numeric values as numbers instead of
strings.
* The `mon_osd_max_op_age` option has been renamed to
`mon_osd_warn_op_age` (default: 32 seconds), to indicate we
generate a warning at this age. There is also a new
`mon_osd_err_op_age_ratio` that is a expressed as a multitple of
`mon_osd_warn_op_age` (default: 128, for roughly 60 minutes) to
control when an error is generated.
* The default maximum size for a single RADOS object has been reduced from
100GB to 128MB. The 100GB limit was completely impractical in practice
while the 128MB limit is a bit high but not unreasonable. If you have an
application written directly to librados that is using objects larger than
128MB you may need to adjust `osd_max_object_size`.
* The semantics of the 'rados ls' and librados object listing
operations have always been a bit confusing in that "whiteout"
objects (which logically don't exist and will return ENOENT if you
try to access them) are included in the results. Previously
whiteouts only occurred in cache tier pools. In luminous, logically
deleted but snapshotted objects now result in a whiteout object, and
as a result they will appear in 'rados ls' results, even though
trying to read such an object will result in ENOENT. The 'rados
listsnaps' operation can be used in such a case to enumerate which
snapshots are present.
This may seem a bit strange, but is less strange than having a
deleted-but-snapshotted object not appear at all and be completely
hidden from librados's ability to enumerate objects. Future
versions of Ceph will likely include an alternative object
enumeration interface that makes it more natural and efficient to
enumerate all objects along with their snapshot and clone metadata.
* *ceph-mgr*:
- There is a new daemon, *ceph-mgr*, which is a required part of any
Ceph deployment. Although IO can continue when *ceph-mgr* is
down, metrics will not refresh and some metrics-related calls
(e.g., `ceph df`) may block. We recommend deploying several instances of
*ceph-mgr* for reliability. See the notes on `Upgrading`_ below.
- The *ceph-mgr* daemon includes a REST-based management API. The
API is still experimental and somewhat limited but will form the basis
for API-based management of Ceph going forward.
- The `status` ceph-mgr module is enabled by default, and initially provides two
commands: `ceph tell mgr osd status` and `ceph tell mgr fs status`. These
are high level colorized views to complement the existing CLI.
* The overall *scalability* of the cluster has improved. We have
successfully tested clusters with up to 10,000 OSDs.
* Each OSD can now have a *device class* associated with it (e.g., `hdd` or
`ssd`), allowing CRUSH rules to trivially map data to a subset of devices
in the system. Manually writing CRUSH rules or manual editing of the CRUSH
is normally not required.
* You can now *optimize CRUSH weights* can now be optimized to
maintain a *near-perfect distribution of data* across OSDs.
* There is also a new `upmap` exception mechanism that allows
individual PGs to be moved around to achieve a *perfect
distribution* (this requires luminous clients).
* Each OSD now adjusts its default configuration based on whether the
backing device is an HDD or SSD. Manual tuning generally not required.
* The prototype *mclock QoS queueing algorithm* is now available.
* There is now a *backoff* mechanism that prevents OSDs from being
overloaded by requests to objects or PGs that are not currently able to
process IO.
* There is a *simplified OSD replacement process* that is more robust.
* You can query the supported features and (apparent) releases of
all connected daemons and clients with `ceph features`.
* You can configure the oldest Ceph client version you wish to allow to
connect to the cluster via `ceph osd set-require-min-compat-client` and
Ceph will prevent you from enabling features that will break compatibility
with those clients.
* Several `sleep` settings, include `osd_recovery_sleep`,
`osd_snap_trim_sleep`, and `osd_scrub_sleep` have been
reimplemented to work efficiently. (These are used in some cases
to work around issues throttling background work.)
* The deprecated 'crush_ruleset' property has finally been removed; please use
'crush_rule' instead for the 'osd pool get ...' and 'osd pool set ..' commands.
* The 'osd pool default crush replicated ruleset' option has been
removed and replaced by the 'osd pool default crush rule' option.
By default it is -1, which means the mon will pick the first type
replicated rule in the CRUSH map for replicated pools. Erasure
coded pools have rules that are automatically created for them if they are
not specified at pool creation time.
- *RGW*:
* RGW *metadata search* backed by ElasticSearch now supports end
user requests service via RGW itself, and also supports custom
metadata fields. A query language a set of RESTful APIs were
created for users to be able to search objects by their
metadata. New APIs that allow control of custom metadata fields
were also added.
* RGW now supports *dynamic bucket index sharding*. As the number
of objects in a bucket grows, RGW will automatically reshard the
bucket index in response. No user intervention or bucket size
capacity planning is required.
* RGW introduces *server side encryption* of uploaded objects with
three options for the management of encryption keys: automatic
encryption (only recommended for test setups), customer provided
keys similar to Amazon SSE-C specification, and through the use of
an external key management service (Openstack Barbican) similar
to Amazon SSE-KMS specification.
* RGW now has preliminary AWS-like bucket policy API support. For
now, policy is a means to express a range of new authorization
concepts. In the future it will be the foundation for additional
auth capabilities such as STS and group policy.
* RGW has consolidated the several metadata index pools via the use of rados
namespaces.
- *RBD*:
* RBD now has full, stable support for *erasure coded pools* via the new
`--data-pool` option to `rbd create`.
* RBD mirroring's rbd-mirror daemon is now highly available. We
recommend deploying several instances of rbd-mirror for
reliability.
* The default 'rbd' pool is no longer created automatically during
cluster creation. Additionally, the name of the default pool used
by the rbd CLI when no pool is specified can be overridden via a
new `rbd default pool = <pool name>` configuration option.
* Initial support for deferred image deletion via new `rbd
trash` CLI commands. Images, even ones actively in-use by
clones, can be moved to the trash and deleted at a later time.
* New pool-level `rbd mirror pool promote` and `rbd mirror pool
demote` commands to batch promote/demote all mirrored images
within a pool.
* Mirroring now optionally supports a configurable replication delay
via the `rbd mirroring replay delay = <seconds>` configuration
option.
* Improved discard handling when the object map feature is enabled.
* rbd CLI `import` and `copy` commands now detect sparse and
preserve sparse regions.
* Images and Snapshots will now include a creation timestamp
- *CephFS*:
* *Multiple active MDS daemons* is now considered stable. The number
of active MDS servers may be adjusted up or down on an active CephFS file
system.
* CephFS *directory fragmentation* is now stable and enabled by
default on new filesystems. To enable it on existing filesystems
use "ceph fs set <fs_name> allow_dirfrags". Large or very busy
directories are sharded and (potentially) distributed across
multiple MDS daemons automatically.
* Directory subtrees can be explicitly pinned to specific MDS daemons in
cases where the automatic load balancing is not desired or effective.
- *Miscellaneous*:
* Release packages are now being built for *Debian Stretch*. The
distributions we build for now includes:
- CentOS 7 (x86_64 and aarch64)
- Debian 8 Jessie (x86_64)
- Debian 9 Stretch (x86_64)
- Ubuntu 16.04 Xenial (x86_64 and aarch64)
- Ubuntu 14.04 Trusty (x86_64)
Note that QA is limited to CentOS and Ubuntu (xenial and trusty).
* *CLI changes*:
- The `ceph -s` or `ceph status` command has a fresh look.
- `ceph {osd,mds,mon} versions` summarizes versions of running daemons.
- `ceph {osd,mds,mon} count-metadata <property>` similarly
tabulates any other daemon metadata visible via the `ceph
{osd,mds,mon} metadata` commands.
- `ceph features` summarizes features and releases of connected
clients and daemons.
- `ceph osd require-osd-release <release>` replaces the old
`require_RELEASE_osds` flags.
- `ceph osd pg-upmap`, `ceph osd rm-pg-upmap`, `ceph osd
pg-upmap-items`, `ceph osd rm-pg-upmap-items` can explicitly
manage `upmap` items.
- `ceph osd getcrushmap` returns a crush map version number on
stderr, and `ceph osd setcrushmap [version]` will only inject
an updated crush map if the version matches. This allows crush
maps to be updated offline and then reinjected into the cluster
without fear of clobbering racing changes (e.g., by newly added
osds or changes by other administrators).
- `ceph osd create` has been replaced by `ceph osd new`. This
should be hidden from most users by user-facing tools like
`ceph-disk`.
- `ceph osd destroy` will mark an OSD destroyed and remove its
cephx and lockbox keys. However, the OSD id and CRUSH map entry
will remain in place, allowing the id to be reused by a
replacement device with minimal data rebalancing.
- `ceph osd purge` will remove all traces of an OSD from the
cluster, including its cephx encryption keys, dm-crypt lockbox
keys, OSD id, and crush map entry.
- `ceph osd ls-tree <name>` will output a list of OSD ids under
the given CRUSH name (like a host or rack name). This is useful
for applying changes to entire subtrees. For example, `ceph
osd down `ceph osd ls-tree rack1``.
- `ceph osd {add,rm}-{noout,noin,nodown,noup}` allow the
`noout`, `nodown`, `noin`, and `noup` flags to be applied to
specific OSDs.
- `ceph log last [n]` will output the last *n* lines of the cluster
log.
- `ceph mgr dump` will dump the MgrMap, including the currently active
ceph-mgr daemon and any standbys.
- `ceph mgr module ls` will list active ceph-mgr modules.
- `ceph mgr module {enable,disable} <name>` will enable or
disable the named mgr module. The module must be present in the
configured `mgr_module_path` on the host(s) where `ceph-mgr` is
running.
- `ceph osd crush swap-bucket <src> <dest>` will swap the
contents of two CRUSH buckets in the hierarchy while preserving
the buckets' ids. This allows an entire subtree of devices to
be replaced (e.g., to replace an entire host of FileStore OSDs
with newly-imaged BlueStore OSDs) without disrupting the
distribution of data across neighboring devices.
- `ceph osd set-require-min-compat-client <release>` configures
the oldest client release the cluster is required to support.
Other changes, like CRUSH tunables, will fail with an error if
they would violate this setting. Changing this setting also
fails if clients older than the specified release are currently
connected to the cluster.
- `ceph config-key dump` dumps config-key entries and their
contents. (The existing `ceph config-key list` only dumps the key
names, not the values.)
- `ceph osd set-{full,nearfull,backfillfull}-ratio` sets the
cluster-wide ratio for various full thresholds (when the cluster
refuses IO, when the cluster warns about being close to full,
when an OSD will defer rebalancing a PG to itself,
respectively).
- `ceph osd reweightn` will specify the `reweight` values for
multiple OSDs in a single command. This is equivalent to a series of
`ceph osd reweight` commands.
- `ceph osd crush class {create,rm,ls,rename}` manage the new
CRUSH *device class* feature. `ceph crush set-device-class
<class> <osd> [<osd>...]` will set the class for particular devices.
- `ceph osd crush rule create-replicated` replaces the old
`ceph osd crush rule create-simple` command to create a CRUSH
rule for a replicated pool. Notably it takes a `class` argument
for the *device class* the rule should target (e.g., `ssd` or
`hdd`).
- `ceph mon feature ls` will list monitor features recorded in the
MonMap. `ceph mon feature set` will set an optional feature (none of
these exist yet).
- `ceph tell <daemon> help` will now return a usage summary.
Major Changes from Jewel
------------------------
- *RADOS*:
* We now default to the AsyncMessenger (`ms type = async`) instead
of the legacy SimpleMessenger. The most noticeable difference is
that we now use a fixed sized thread pool for network connections
(instead of two threads per socket with SimpleMessenger).
* Some OSD failures are now detected almost immediately, whereas
previously the heartbeat timeout (which defaults to 20 seconds)
had to expire. This prevents IO from blocking for an extended
period for failures where the host remains up but the ceph-osd
process is no longer running.
* The size of encoded OSDMaps has been reduced.
* The OSDs now quiesce scrubbing when recovery or rebalancing is in progress.
- *RGW*:
* RGW now supports the S3 multipart object copy-part API.
* It is possible now to reshard an existing bucket offline. Offline
bucket resharding currently requires that all IO (especially
writes) to the specific bucket is quiesced. (For automatic online
resharding, see the new feature in Luminous above.)
* RGW now supports data compression for objects.
* Civetweb version has been upgraded to 1.8
* The Swift static website API is now supported (S3 support has been added
previously).
* S3 bucket lifecycle API has been added. Note that currently it only supports
object expiration.
* Support for custom search filters has been added to the LDAP auth
implementation.
* Support for NFS version 3 has been added to the RGW NFS gateway.
* A Python binding has been created for librgw.
- *RBD*:
* The rbd-mirror daemon now supports replicating dynamic image
feature updates and image metadata key/value pairs from the
primary image to the non-primary image.
* The number of image snapshots can be optionally restricted to a
configurable maximum.
* The rbd Python API now supports asynchronous IO operations.
- *CephFS*:
* libcephfs function definitions have been changed to enable proper
uid/gid control. The library version has been increased to reflect the
interface change.
* Standby replay MDS daemons now consume less memory on workloads
doing deletions.
* Scrub now repairs backtrace, and populates `damage ls` with
discovered errors.
* A new `pg_files` subcommand to `cephfs-data-scan` can identify
files affected by a damaged or lost RADOS PG.
* The false-positive "failing to respond to cache pressure" warnings have
been fixed.
Upgrade from Jewel or Kraken
----------------------------
.. _Upgrading:
#. Ensure that the `sortbitwise` flag is enabled::
# ceph osd set sortbitwise
#. Make sure your cluster is stable and healthy (no down or
recoverying OSDs). (Optional, but recommended.)
#. Do not create any new erasure-code pools while upgrading the monitors.
#. Set the `noout` flag for the duration of the upgrade. (Optional
but recommended.)::
# ceph osd set noout
#. Upgrade monitors by installing the new packages and restarting the
monitor daemons. Note that, unlike prior releases, the ceph-mon
daemons *must* be upgraded first.::
# systemctl restart ceph-mon.target
Verify the monitor upgrade is complete once all monitors are up by
looking for the `luminous` feature string in the mon map. For
example::
# ceph mon feature ls
should include `luminous` under persistent features::
on current monmap (epoch NNN)
persistent: [kraken,luminous]
required: [kraken,luminous]
#. Add or restart `ceph-mgr` daemons. If you are upgrading from
kraken, upgrade packages and restart ceph-mgr daemons with::
# systemctl restart ceph-mgr.target
If you are upgrading from kraken, you may already have ceph-mgr
daemons deployed. If not, or if you are upgrading from jewel, you
can deploy new daemons with tools like ceph-deploy or ceph-ansible.
For example,::
# ceph-deploy mgr create HOST
Verify the ceph-mgr daemons are running by checking `ceph -s`::
# ceph -s
...
services:
mon: 3 daemons, quorum foo,bar,baz
mgr: foo(active), standbys: bar, baz
...
#. Upgrade all OSDs by installing the new packages and restarting the
ceph-osd daemons on all hosts.::
# systemctl restart ceph-osd.target
You can monitor the progress of the OSD upgrades with the new
`ceph osd versions` command.::
# ceph osd versions
{
"ceph version 12.2.0 (...) luminous (stable)": 12,
"ceph version 10.2.6 (...)": 3,
}
#. Upgrade all CephFS daemons by upgrading packages and restarting
daemons on all hosts.::
# systemctl restart ceph-mds.target
#. Upgrade all radosgw daemons by upgrading packages and restarting
daemons on all hosts.::
# systemctl restart radosgw.target
#. Complete the upgrade by disallowing pre-luminous OSDs::
# ceph osd require-osd-release luminous
If you set `noout` at the beginning, be sure to clear it with::
# ceph osd unset noout
#. Verify the cluster is healthy with `ceph health`.
Upgrading from pre-Jewel releases (like Hammer)
-----------------------------------------------
You *must* first upgrade to Jewel (10.2.z) before attempting an
upgrade to Luminous.
Upgrade compatibility notes, Kraken to Luminous
-----------------------------------------------
* We no longer test the FileStore ceph-osd backend in combination with
`btrfs`. We recommend against using btrfs. If you are using
btrfs-based OSDs and want to upgrade to luminous you will need to
add the follwing to your ceph.conf::
enable experimental unrecoverable data corrupting features = btrfs
The code is mature and unlikely to change, but we are only
continuing to test the Jewel stable branch against btrfs. We
recommend moving these OSDs to FileStore with XFS or BlueStore.
* The `ruleset-*` properties for the erasure code profiles have been
renamed to `crush-*` to (1) move away from the obsolete 'ruleset'
term and to be more clear about their purpose. There is also a new
optional `crush-device-class` property to specify a CRUSH device
class to use for the erasure coded pool. Existing erasure code
profiles will be converted automatically when upgrade completes
(when the `ceph osd require-osd-release luminous` command is run)
but any provisioning tools that create erasure coded pools may need
to be updated.
* When assigning a network to the public network and not to
the cluster network the network specification of the public
network will be used for the cluster network as well.
In older versions this would lead to cluster services
being bound to 0.0.0.0:<port>, thus making the
cluster service even more publicly available than the
public services. When only specifying a cluster network it
will still result in the public services binding to 0.0.0.0.
* In previous versions, if a client sent an op to the wrong OSD, the OSD
would reply with ENXIO. The rationale here is that the client or OSD is
clearly buggy and we want to surface the error as clearly as possible.
We now only send the ENXIO reply if the osd_enxio_on_misdirected_op option
is enabled (it's off by default). This means that a VM using librbd that
previously would have gotten an EIO and gone read-only will now see a
blocked/hung IO instead.
* The "journaler allow split entries" config setting has been removed.
- *librados*:
* Some variants of the omap_get_keys and omap_get_vals librados
functions have been deprecated in favor of omap_get_vals2 and
omap_get_keys2. The new methods include an output argument
indicating whether there are additional keys left to fetch.
Previously this had to be inferred from the requested key count vs
the number of keys returned, but this breaks with new OSD-side
limits on the number of keys or bytes that can be returned by a
single omap request. These limits were introduced by kraken but
are effectively disabled by default (by setting a very large limit
of 1 GB) because users of the newly deprecated interface cannot
tell whether they should fetch more keys or not. In the case of
the standalone calls in the C++ interface
(IoCtx::get_omap_{keys,vals}), librados has been updated to loop on
the client side to provide a correct result via multiple calls to
the OSD. In the case of the methods used for building
multi-operation transactions, however, client-side looping is not
practical, and the methods have been deprecated. Note that use of
either the IoCtx methods on older librados versions or the
deprecated methods on any version of librados will lead to
incomplete results if/when the new OSD limits are enabled.
* The original librados rados_objects_list_open (C) and objects_begin
(C++) object listing API, deprecated in Hammer, has finally been
removed. Users of this interface must update their software to use
either the rados_nobjects_list_open (C) and nobjects_begin (C++) API or
the new rados_object_list_begin (C) and object_list_begin (C++) API
before updating the client-side librados library to Luminous.
Object enumeration (via any API) with the latest librados version
and pre-Hammer OSDs is no longer supported. Note that no in-tree
Ceph services rely on object enumeration via the deprecated APIs, so
only external librados users might be affected.
The newest (and recommended) rados_object_list_begin (C) and
object_list_begin (C++) API is only usable on clusters with the
SORTBITWISE flag enabled (Jewel and later). (Note that this flag is
required to be set before upgrading beyond Jewel.)
- *CephFS*:
* When configuring ceph-fuse mounts in /etc/fstab, a new syntax is
available that uses "ceph.<arg>=<val>" in the options column, instead
of putting configuration in the device column. The old style syntax
still works. See the documentation page "Mount CephFS in your
file systems table" for details.
* CephFS clients without the 'p' flag in their authentication capability
string will no longer be able to set quotas or any layout fields. This
flag previously only restricted modification of the pool and namespace
fields in layouts.
* CephFS will generate a health warning if you have fewer standby daemons
than it thinks you wanted. By default this will be 1 if you ever had
a standby, and 0 if you did not. You can customize this using
`ceph fs set <fs> standby_count_wanted <number>`. Setting it
to zero will effectively disable the health check.
* The "ceph mds tell ..." command has been removed. It is superceded
by "ceph tell mds.<id> ..."
Notable Changes since v12.1.1 (RC1)
-----------------------------------
* choose_args encoding has been changed to make it architecture-independent.
If you deployed Luminous dev releases or 12.1.0 rc release and made use of
the CRUSH choose_args feature, you need to remove all choose_args mappings
from your CRUSH map before starting the upgrade.
* The 'ceph health' structured output (JSON or XML) no longer contains
a 'timechecks' section describing the time sync status. This
information is now available via the 'ceph time-sync-status'
command.
* Certain extra fields in the 'ceph health' structured output that
used to appear if the mons were low on disk space (which duplicated
the information in the normal health warning messages) are now gone.
* The "ceph -w" output no longer contains audit log entries by default.
Add a "--watch-channel=audit" or "--watch-channel=*" to see them.
* The 'apply' mode of cephfs-journal-tool has been removed
* Added new configuration "public bind addr" to support dynamic environments
like Kubernetes. When set the Ceph MON daemon could bind locally to an IP
address and advertise a different IP address "public addr" on the network.
For a detailed changelog refer to the blog post entry at
http://ceph.com/releases/v12-1-1-luminous-rc-released/
Getting Ceph
------------
* Git at git://github.com/ceph/ceph.git
* Tarball at http://download.ceph.com/tarballs/ceph-12.1.1.tar.gz
* For packages, see http://docs.ceph.com/docs/master/install/get-packages/
* For ceph-deploy, see http://docs.ceph.com/docs/master/install/install-ceph-deploy
* Release sha1: f3e663a190bf2ed12c7e3cda288b9a159572c800
--
Abhishek Lekshmanan
SUSE Linux GmbH, GF: Felix Imendörffer, Jane Smithard, Graham Norton,
HRB 21284 (AG Nürnberg)
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
