ceph/doc/rados/operations/monitoring-osd-pg.rst

618 lines
24 KiB
ReStructuredText
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

=========================
Monitoring OSDs and PGs
=========================
High availability and high reliability require a fault-tolerant approach to
managing hardware and software issues. Ceph has no single point-of-failure, and
can service requests for data in a "degraded" mode. Ceph's `data placement`_
introduces a layer of indirection to ensure that data doesn't bind directly to
particular OSD addresses. This means that tracking down system faults requires
finding the `placement group`_ and the underlying OSDs at root of the problem.
.. tip:: A fault in one part of the cluster may prevent you from accessing a
particular object, but that doesn't mean that you can't access other objects.
When you run into a fault, don't panic. Just follow the steps for monitoring
your OSDs and placement groups. Then, begin troubleshooting.
Ceph is generally self-repairing. However, when problems persist, monitoring
OSDs and placement groups will help you identify the problem.
Monitoring OSDs
===============
An OSD's status is either in the cluster (``in``) or out of the cluster
(``out``); and, it is either up and running (``up``), or it is down and not
running (``down``). If an OSD is ``up``, it may be either ``in`` the cluster
(you can read and write data) or it is ``out`` of the cluster. If it was
``in`` the cluster and recently moved ``out`` of the cluster, Ceph will migrate
placement groups to other OSDs. If an OSD is ``out`` of the cluster, CRUSH will
not assign placement groups to the OSD. If an OSD is ``down``, it should also be
``out``.
.. note:: If an OSD is ``down`` and ``in``, there is a problem and the cluster
will not be in a healthy state.
.. ditaa:: +----------------+ +----------------+
| | | |
| OSD #n In | | OSD #n Up |
| | | |
+----------------+ +----------------+
^ ^
| |
| |
v v
+----------------+ +----------------+
| | | |
| OSD #n Out | | OSD #n Down |
| | | |
+----------------+ +----------------+
If you execute a command such as ``ceph health``, ``ceph -s`` or ``ceph -w``,
you may notice that the cluster does not always echo back ``HEALTH OK``. Don't
panic. With respect to OSDs, you should expect that the cluster will **NOT**
echo ``HEALTH OK`` in a few expected circumstances:
#. You haven't started the cluster yet (it won't respond).
#. You have just started or restarted the cluster and it's not ready yet,
because the placement groups are getting created and the OSDs are in
the process of peering.
#. You just added or removed an OSD.
#. You just have modified your cluster map.
An important aspect of monitoring OSDs is to ensure that when the cluster
is up and running that all OSDs that are ``in`` the cluster are ``up`` and
running, too. To see if all OSDs are running, execute::
ceph osd stat
The result should tell you the map epoch (eNNNN), the total number of OSDs (x),
how many are ``up`` (y) and how many are ``in`` (z). ::
eNNNN: x osds: y up, z in
If the number of OSDs that are ``in`` the cluster is more than the number of
OSDs that are ``up``, execute the following command to identify the ``ceph-osd``
daemons that aren't running::
ceph osd tree
::
dumped osdmap tree epoch 1
# id weight type name up/down reweight
-1 2 pool openstack
-3 2 rack dell-2950-rack-A
-2 2 host dell-2950-A1
0 1 osd.0 up 1
1 1 osd.1 down 1
.. tip:: The ability to search through a well-designed CRUSH hierarchy may help
you troubleshoot your cluster by identifying the physcial locations faster.
If an OSD is ``down``, start it::
sudo /etc/init.d/ceph -a start osd.1
See `OSD Not Running`_ for problems associated with OSDs that stopped, or won't
restart.
PG Sets
=======
When CRUSH assigns placement groups to OSDs, it looks at the number of replicas
for the pool and assigns the placement group to OSDs such that each replica of
the placement group gets assigned to a different OSD. For example, if the pool
requires three replicas of a placement group, CRUSH may assign them to
``osd.1``, ``osd.2`` and ``osd.3`` respectively. CRUSH actually seeks a
pseudo-random placement that will take into account failure domains you set in
your `CRUSH map`_, so you will rarely see placement groups assigned to nearest
neighbor OSDs in a large cluster. We refer to the set of OSDs that should
contain the replicas of a particular placement group as the **Acting Set**. In
some cases, an OSD in the Acting Set is ``down`` or otherwise not able to
service requests for objects in the placement group. When these situations
arise, don't panic. Common examples include:
- You added or removed an OSD. Then, CRUSH reassigned the placement group to
other OSDs--thereby changing the composition of the Acting Set and spawning
the migration of data with a "backfill" process.
- An OSD was ``down``, was restared, and is now ``recovering``.
- An OSD in the Acting Set is ``down`` or unable to service requests,
and another OSD has temporarily assumed its duties.
Ceph processes a client request using the **Up Set**, which is the set of OSDs
that will actually handle the requests. In most cases, the Up Set and the Acting
Set are virtually identical. When they are not, it may indicate that Ceph is
migrating data, an OSD is recovering, or that there is a problem (i.e., Ceph
usually echoes a "HEALTH WARN" state with a "stuck stale" message in such
scenarios).
To retrieve a list of placement groups, execute::
ceph pg dump
To view which OSDs are within the Acting Set or the Up Set for a given placement
group, execute::
ceph pg map {pg-num}
The result should tell you the osdmap epoch (eNNN), the placement group number
({pg-num}), the OSDs in the Up Set (up[]), and the OSDs in the acting set
(acting[]). ::
osdmap eNNN pg {pg-num} -> up [0,1,2] acting [0,1,2]
.. note:: If the Up Set and Acting Set do not match, this may be an indicator
that the cluster rebalancing itself or of a potential problem with
the cluster.
Peering
=======
Before you can write data to a placement group, it must be in an ``active``
state, and it **should** be in a ``clean`` state. For Ceph to determine the
current state of a placement group, the primary OSD of the placement group
(i.e., the first OSD in the acting set), peers with the secondary and tertiary
OSDs to establish agreement on the current state of the placement group
(assuming a pool with 3 replicas of the PG).
.. ditaa:: +---------+ +---------+ +-------+
| OSD 1 | | OSD 2 | | OSD 3 |
+---------+ +---------+ +-------+
| | |
| Request To | |
| Peer | |
|-------------->| |
|<--------------| |
| Peering |
| |
| Request To |
| Peer |
|----------------------------->|
|<-----------------------------|
| Peering |
The OSDs also report their status to the monitor. See `Configuring Monitor/OSD
Interaction`_ for details. To troubleshoot peering issues, see `Peering
Failure`_.
Monitoring Placement Group States
=================================
If you execute a command such as ``ceph health``, ``ceph -s`` or ``ceph -w``,
you may notice that the cluster does not always echo back ``HEALTH OK``. After
you check to see if the OSDs are running, you should also check placement group
states. You should expect that the cluster will **NOT** echo ``HEALTH OK`` in a
number of placement group peering-related circumstances:
#. You have just created a pool and placement groups haven't peered yet.
#. The placement groups are recovering.
#. You have just added an OSD to or removed an OSD from the cluster.
#. You have just modified your CRUSH map and your placement groups are migrating.
#. There is inconsistent data in different replicas of a placement group.
#. Ceph is scrubbing a placement group's replicas.
#. Ceph doesn't have enough storage capacity to complete backfilling operations.
If one of the foregoing circumstances causes Ceph to echo ``HEALTH WARN``, don't
panic. In many cases, the cluster will recover on its own. In some cases, you
may need to take action. An important aspect of monitoring placement groups is
to ensure that when the cluster is up and running that all placement groups are
``active``, and preferably in the ``clean`` state. To see the status of all
placement groups, execute::
ceph pg stat
The result should tell you the placement group map version (vNNNNNN), the total
number of placement groups (x), and how many placement groups are in a
particular state such as ``active+clean`` (y). ::
vNNNNNN: x pgs: y active+clean; z bytes data, aa MB used, bb GB / cc GB avail
.. note:: It is common for Ceph to report multiple states for placement groups.
In addition to the placement group states, Ceph will also echo back the amount
of data used (aa), the amount of storage capacity remaining (bb), and the total
storage capacity for the placement group. These numbers can be important in a
few cases:
- You are reaching your ``near full ratio`` or ``full ratio``.
- Your data isn't getting distributed across the cluster due to an
error in your CRUSH configuration.
.. topic:: Placement Group IDs
Placement group IDs consist of the pool number (not pool name) followed
by a period (.) and the placement group ID--a hexadecimal number. You
can view pool numbers and their names from the output of ``ceph osd
lspools``. The default pool names ``data``, ``metadata`` and ``rbd``
correspond to pool numbers ``0``, ``1`` and ``2`` respectively. A fully
qualified placement group ID has the following form::
{pool-num}.{pg-id}
And it typically looks like this::
0.1f
To retrieve a list of placement groups, execute the following::
ceph pg dump
You can also format the output in JSON format and save it to a file::
ceph pg dump -o {filename} --format=json
To query a particular placement group, execute the following::
ceph pg {poolnum}.{pg-id} query
Ceph will output the query in JSON format.
.. code-block:: javascript
{
"state": "active+clean",
"up": [
1,
0
],
"acting": [
1,
0
],
"info": {
"pgid": "1.e",
"last_update": "4'1",
"last_complete": "4'1",
"log_tail": "0'0",
"last_backfill": "MAX",
"purged_snaps": "[]",
"history": {
"epoch_created": 1,
"last_epoch_started": 537,
"last_epoch_clean": 537,
"last_epoch_split": 534,
"same_up_since": 536,
"same_interval_since": 536,
"same_primary_since": 536,
"last_scrub": "4'1",
"last_scrub_stamp": "2013-01-25 10:12:23.828174"
},
"stats": {
"version": "4'1",
"reported": "536'782",
"state": "active+clean",
"last_fresh": "2013-01-25 10:12:23.828271",
"last_change": "2013-01-25 10:12:23.828271",
"last_active": "2013-01-25 10:12:23.828271",
"last_clean": "2013-01-25 10:12:23.828271",
"last_unstale": "2013-01-25 10:12:23.828271",
"mapping_epoch": 535,
"log_start": "0'0",
"ondisk_log_start": "0'0",
"created": 1,
"last_epoch_clean": 1,
"parent": "0.0",
"parent_split_bits": 0,
"last_scrub": "4'1",
"last_scrub_stamp": "2013-01-25 10:12:23.828174",
"log_size": 128,
"ondisk_log_size": 128,
"stat_sum": {
"num_bytes": 205,
"num_objects": 1,
"num_object_clones": 0,
"num_object_copies": 0,
"num_objects_missing_on_primary": 0,
"num_objects_degraded": 0,
"num_objects_unfound": 0,
"num_read": 1,
"num_read_kb": 0,
"num_write": 3,
"num_write_kb": 1
},
"stat_cat_sum": {
},
"up": [
1,
0
],
"acting": [
1,
0
]
},
"empty": 0,
"dne": 0,
"incomplete": 0
},
"recovery_state": [
{
"name": "Started\/Primary\/Active",
"enter_time": "2013-01-23 09:35:37.594691",
"might_have_unfound": [
],
"scrub": {
"scrub_epoch_start": "536",
"scrub_active": 0,
"scrub_block_writes": 0,
"finalizing_scrub": 0,
"scrub_waiting_on": 0,
"scrub_waiting_on_whom": [
]
}
},
{
"name": "Started",
"enter_time": "2013-01-23 09:35:31.581160"
}
]
}
The following subsections describe common states in greater detail.
Creating
--------
When you create a pool, it will create the number of placement groups you
specified. Ceph will echo ``creating`` when it is creating one or more
placement groups. Once they are created, the OSDs that are part of a placement
group's Acting Set will peer. Once peering is complete, the placement group
status should be ``active+clean``, which means a Ceph client can begin writing
to the placement group.
.. ditaa::
/-----------\ /-----------\ /-----------\
| Creating |------>| Peering |------>| Active |
\-----------/ \-----------/ \-----------/
Peering
-------
When Ceph is Peering a placement group, Ceph is bringing the OSDs that
store the replicas of the placement group into **agreement about the state**
of the objects and metadata in the placement group. When Ceph completes peering,
this means that the OSDs that store the placement group agree about the current
state of the placement group. However, completion of the peering process does
**NOT** mean that each replica has the latest contents.
.. topic:: Authoratative History
Ceph will **NOT** acknowledge a write operation to a client, until
all OSDs of the acting set persist the write operation. This practice
ensures that at least one member of the acting set will have a record
of every acknowledged write operation since the last successful
peering operation.
With an accurate record of each acknowledged write operation, Ceph can
construct and disseminate a new authoritative history of the placement
group--a complete, and fully ordered set of operations that, if performed,
would bring an OSDs copy of a placement group up to date.
Active
------
Once Ceph completes the peering process, a placement group may become
``active``. The ``active`` state means that the data in the placement group is
generally available in the primary placement group and the replicas for read
and write operations.
Clean
-----
When a placement group is in the ``clean`` state, the primary OSD and the
replica OSDs have successfully peered and there are no stray replicas for the
placement group. Ceph replicated all objects in the placement group the correct
number of times.
Degraded
--------
When a client writes an object to the primary OSD, the primary OSD is
responsible for writing the replicas to the replica OSDs. After the primary OSD
writes the object to storage, the placement group will remain in a ``degraded``
state until the primary OSD has received an acknowledgement from the replica
OSDs that Ceph created the replica objects successfully.
The reason a placement group can be ``active+degraded`` is that an OSD may be
``active`` even though it doesn't hold all of the objects yet. If an OSD goes
``down``, Ceph marks each placement group assigned to the OSD as ``degraded``.
The OSDs must peer again when the OSD comes back online. However, a client can
still write a new object to a ``degraded`` placement group if it is ``active``.
If an OSD is ``down`` and the ``degraded`` condition persists, Ceph may mark the
``down`` OSD as ``out`` of the cluster and remap the data from the ``down`` OSD
to another OSD. The time between being marked ``down`` and being marked ``out``
is controlled by ``mon osd down out interval``, which is set to ``300`` seconds
by default.
A placement group can also be ``degraded``, because Ceph cannot find one or more
objects that Ceph thinks should be in the placement group. While you cannot
read or write to unfound objects, you can still access all of the other objects
in the ``degraded`` placement group.
Recovering
----------
Ceph was designed for fault-tolerance at a scale where hardware and software
problems are ongoing. When an OSD goes ``down``, its contents may fall behind
the current state of other replicas in the placement groups. When the OSD is
back ``up``, the contents of the placement groups must be updated to reflect the
current state. During that time period, the OSD may reflect a ``recovering``
state.
Recovery isn't always trivial, because a hardware failure might cause a
cascading failure of multiple OSDs. For example, a network switch for a rack or
cabinet may fail, which can cause the OSDs of a number of host machines to fall
behind the current state of the cluster. Each one of the OSDs must recover once
the fault is resolved.
Ceph provides a number of settings to balance the resource contention between
new service requests and the need to recover data objects and restore the
placement groups to the current state. The ``osd recovery delay start`` setting
allows an OSD to restart, re-peer and even process some replay requests before
starting the recovery process. The ``osd recovery threads`` setting limits the
number of threads for the recovery process (1 thread by default). The ``osd
recovery thread timeout`` sets a thread timeout, because multiple OSDs may fail,
restart and re-peer at staggered rates. The ``osd recovery max active`` setting
limits the number of recovery requests an OSD will entertain simultaneously to
prevent the OSD from failing to serve . The ``osd recovery max chunk`` setting
limits the size of the recovered data chunks to prevent network congestion.
Back Filling
------------
When a new OSD joins the cluster, CRUSH will reassign placement groups from OSDs
in the cluster to the newly added OSD. Forcing the new OSD to accept the
reassigned placement groups immediately can put excessive load on the new OSD.
Back filling the OSD with the placement groups allows this process to begin in
the background. Once backfilling is complete, the new OSD will begin serving
requests when it is ready.
During the backfill operations, you may see one of several states:
``backfill_wait`` indicates that a backfill operation is pending, but isn't
underway yet; ``backfill`` indicates that a backfill operation is underway;
and, ``backfill_too_full`` indicates that a backfill operation was requested,
but couldn't be completed due to insufficient storage capacity. When a
placement group can't be backfilled, it may be considered ``incomplete``.
Ceph provides a number of settings to manage the load spike associated with
reassigning placement groups to an OSD (especially a new OSD). By default,
``osd_max_backfills`` sets the maximum number of concurrent backfills to or from
an OSD to 10. The ``osd backfill full ratio`` enables an OSD to refuse a
backfill request if the OSD is approaching its full ratio (85%, by default).
If an OSD refuses a backfill request, the ``osd backfill retry interval``
enables an OSD to retry the request (after 10 seconds, by default). OSDs can
also set ``osd backfill scan min`` and ``osd backfill scan max`` to manage scan
intervals (64 and 512, by default).
Remapped
--------
When the Acting Set that services a placement group changes, the data migrates
from the old acting set to the new acting set. It may take some time for a new
primary OSD to service requests. So it may ask the old primary to continue to
service requests until the placement group migration is complete. Once data
migration completes, the mapping uses the primary OSD of the new acting set.
Stale
-----
While Ceph uses heartbeats to ensure that hosts and daemons are running, the
``ceph-osd`` daemons may also get into a ``stuck`` state where they aren't
reporting statistics in a timely manner (e.g., a temporary network fault). By
default, OSD daemons report their placement group, up thru, boot and failure
statistics every half second (i.e., ``0.5``), which is more frequent than the
heartbeat thresholds. If the **Primary OSD** of a placement group's acting set
fails to report to the monitor or if other OSDs have reported the primary OSD
``down``, the monitors will mark the placement group ``stale``.
When you start your cluster, it is common to see the ``stale`` state until
the peering process completes. After your cluster has been running for awhile,
seeing placement groups in the ``stale`` state indicates that the primary OSD
for those placement groups is ``down`` or not reporting placement group statistics
to the monitor.
Identifying Troubled PGs
========================
As previously noted, a placement group isn't necessarily problematic just
because its state isn't ``active+clean``. Generally, Ceph's ability to self
repair may not be working when placement groups get stuck. The stuck states
include:
- **Unclean**: Placement groups contain objects that are not replicated the
desired number of times. They should be recovering.
- **Inactive**: Placement groups cannot process reads or writes because they
are waiting for an OSD with the most up-to-date data to come back ``up``.
- **Stale**: Placement groups are in an unknown state, because the OSDs that
host them have not reported to the monitor cluster in a while (configured
by ``mon osd report timeout``).
To identify stuck placement groups, execute the following::
ceph pg dump_stuck [unclean|inactive|stale]
See `Placement Group Subsystem`_ for additional details. To troubleshoot
stuck placement groups, see `Troubleshooting PG Errors`_.
Finding an Object Location
==========================
To store object data in the Ceph Object Store, a Ceph client must:
#. Set an object name
#. Specify a `pool`_
The Ceph client retrieves the latest cluster map and the CRUSH algorithm
calculates how to map the object to a `placement group`_, and then calculates
how to assign the placement group to an OSD dynamically. To find the object
location, all you need is the object name and the pool name. For example::
ceph osd map {poolname} {object-name}
.. topic:: Exercise: Locate an Object
As an exercise, lets create an object. Specify an object name, a path to a
test file containing some object data and a pool name using the
``rados put`` command on the command line. For example::
rados put {object-name} {file-path} --pool=data
rados put test-object-1 testfile.txt --pool=data
To verify that the Ceph Object Store stored the object, execute the following::
rados -p data ls
Now, identify the object location::
ceph osd map {pool-name} {object-name}
ceph osd map data test-object-1
Ceph should output the object's location. For example::
osdmap e537 pool 'data' (0) object 'test-object-1' -> pg 0.d1743484 (0.4) -> up [1,0] acting [1,0]
To remove the test object, simply delete it using the ``rados rm`` command.
For example::
rados rm test-object-1 --pool=data
As the cluster evolves, the object location may change dynamically. One benefit
of Ceph's dynamic rebalancing is that Ceph relieves you from having to perform
the migration manually. See the `Architecture`_ section for details.
.. _data placement: ../data-placement
.. _pool: ../pools
.. _placement group: ../placement-groups
.. _Architecture: ../../../architecture
.. _OSD Not Running: ../../troubleshooting/troubleshooting-osd#osd-not-running
.. _Troubleshooting PG Errors: ../../troubleshooting/troubleshooting-pg#troubleshooting-pg-errors
.. _Peering Failure: ../../troubleshooting/troubleshooting-pg#failures-osd-peering
.. _CRUSH map: ../crush-map
.. _Configuring Monitor/OSD Interaction: ../../configuration/mon-osd-interaction/
.. _Placement Group Subsystem: ../control#placement-group-subsystem