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 it can service requests for data even when in a “degraded” mode. Ceph’s data placement introduces a layer of indirection to ensure that data doesn’t bind directly to specific OSDs. For this reason, tracking system faults requires finding the placement group (PG) and the underlying OSDs at the root of the problem.

Tip

A fault in one part of the cluster might prevent you from accessing a particular object, but that doesn’t mean that you are prevented from accessing other objects. When you run into a fault, don’t panic. Just follow the steps for monitoring your OSDs and placement groups, and then begin troubleshooting.

Ceph is self-repairing. However, when problems persist, monitoring OSDs and placement groups will help you identify the problem.

Monitoring OSDs

An OSD is either in service (in) or out of service (out). An OSD is either running and reachable (up), or it is not running and not reachable (down).

If an OSD is up, it may be either in service (clients can read and write data) or it is out of service. If the OSD was in but then due to a failure or a manual action was set to the out state, Ceph will migrate placement groups to the other OSDs to maintin the configured redundancy.

If an OSD is out of service, CRUSH will not assign placement groups to it. If an OSD is down, it will also be out.

Note

If an OSD is down and in, there is a problem and this indicates that the cluster is not in a healthy state.

If you run the commands ceph health, ceph -s, or ceph -w, you might notice that the cluster does not always show HEALTH OK. Don’t panic. There are certain circumstances in which it is expected and normal that the cluster will NOT show HEALTH OK:

  1. You haven’t started the cluster yet.

  2. You have just started or restarted the cluster and it’s not ready to show health statuses yet, because the PGs are in the process of being created and the OSDs are in the process of peering.

  3. You have just added or removed an OSD.

  4. You have just have modified your cluster map.

Checking to see if OSDs are up and running is an important aspect of monitoring them: whenever the cluster is up and running, every OSD that is in the cluster should also be up and running. To see if all of the cluster’s OSDs are running, run the following command:

ceph osd stat

The output provides the following information: the total number of OSDs (x), how many OSDs are up (y), how many OSDs are in (z), and the map epoch (eNNNN).

x osds: y up, z in; epoch: eNNNN

If the number of OSDs that are in the cluster is greater than the number of OSDs that are up, run the following command to identify the ceph-osd daemons that are not running:

ceph osd tree
#ID CLASS WEIGHT  TYPE NAME             STATUS REWEIGHT PRI-AFF
 -1       2.00000 pool openstack
 -3       2.00000 rack dell-2950-rack-A
 -2       2.00000 host dell-2950-A1
  0   ssd 1.00000      osd.0                up  1.00000 1.00000
  1   ssd 1.00000      osd.1              down  1.00000 1.00000

Tip

Searching through a well-designed CRUSH hierarchy to identify the physical locations of particular OSDs might help you troubleshoot your cluster.

If an OSD is down, start it by running the following command:

sudo systemctl start ceph-osd@1

For problems associated with OSDs that have stopped or won’t restart, see OSD Not Running.

PG Sets

When CRUSH assigns a PG to OSDs, it takes note of how many replicas of the PG are required by the pool and then assigns each replica to a different OSD. For example, if the pool requires three replicas of a PG, CRUSH might assign them individually to osd.1, osd.2 and osd.3. CRUSH seeks a pseudo-random placement that takes into account the failure domains that you have set in your CRUSH map; for this reason, PGs are rarely assigned to immediately adjacent OSDs in a large cluster.

Ceph processes client requests with the Acting Set of OSDs: this is the set of OSDs that currently have a full and working version of a PG shard and that are therefore responsible for handling requests. By contrast, the Up Set is the set of OSDs that contain a shard of a specific PG. Data is moved or copied to the Up Set, or planned to be moved or copied, to the Up Set. See Placement Group Concepts.

Sometimes an OSD in the Acting Set is down or otherwise unable to service requests for objects in the PG. When this kind of situation arises, don’t panic. Common examples of such a situation include:

  • You added or removed an OSD, CRUSH reassigned the PG to other OSDs, and this reassignment changed the composition of the Acting Set and triggered the migration of data by means of a “backfill” process.

  • An OSD was down, was restarted, 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.

Typically, the Up Set and the Acting Set are identical. When they are not, it might indicate that Ceph is migrating the PG (in other words, that the PG has been remapped), that an OSD is recovering, or that there is a problem with the cluster (in such scenarios, Ceph usually shows a “HEALTH WARN” state with a “stuck stale” message).

To retrieve a list of PGs, run the following command:

ceph pg dump

To see which OSDs are within the Acting Set and the Up Set for a specific PG, run the following command:

ceph pg map {pg-num}

The output provides the following information: the osdmap epoch (eNNN), the PG number ({pg-num}), the OSDs in the Up Set (up[]), and the OSDs in the Acting Set (acting[]):

osdmap eNNN pg {raw-pg-num} ({pg-num}) -> up [0,1,2] acting [0,1,2]

Note

If the Up Set and the Acting Set do not match, this might indicate that the cluster is rebalancing itself or that there is a problem with the cluster.

Peering

Before you can write data to a PG, it must be in an active state and it will preferably be in a clean state. For Ceph to determine the current state of a PG, peering must take place. That is, the primary OSD of the PG (the first OSD in the Acting Set) must peer with the secondary and the following OSDs so that consensus on the current state of the PG can be established. In the following diagram, we assume a pool with three replicas of the PG:

The OSDs also report their status to the monitor. For details, see Configuring Monitor/OSD Interaction. To troubleshoot peering issues, see Peering Failure.

Monitoring PG States

If you run the commands ceph health, ceph -s, or ceph -w, you might notice that the cluster does not always show HEALTH OK. After first checking to see if the OSDs are running, you should also check PG states. There are certain PG-peering-related circumstances in which it is expected and normal that the cluster will NOT show HEALTH OK:

  1. You have just created a pool and the PGs haven’t peered yet.

  2. The PGs are recovering.

  3. You have just added an OSD to or removed an OSD from the cluster.

  4. You have just modified your CRUSH map and your PGs are migrating.

  5. There is inconsistent data in different replicas of a PG.

  6. Ceph is scrubbing a PG’s replicas.

  7. Ceph doesn’t have enough storage capacity to complete backfilling operations.

If one of these circumstances causes Ceph to show HEALTH WARN, don’t panic. In many cases, the cluster will recover on its own. In some cases, however, you might need to take action. An important aspect of monitoring PGs is to check their status as active and clean: that is, it is important to ensure that, when the cluster is up and running, all PGs are active and (preferably) clean. To see the status of every PG, run the following command:

ceph pg stat

The output provides the following information: the total number of PGs (x), how many PGs are in a particular state such as active+clean (y), and the amount of data stored (z).

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 PGs (for example, active+clean, active+clean+remapped, active+clean+scrubbing.

Here Ceph shows not only the PG states, but also storage capacity used (aa), the amount of storage capacity remaining (bb), and the total storage capacity of the PG. These values can be important in a few cases:

  • The cluster is reaching its near full ratio or full ratio.

  • Data is not being distributed across the cluster due to an error in the CRUSH configuration.

Placement Group IDs

PG IDs consist of the pool number (not the pool name) followed by a period (.) and a hexadecimal number. You can view pool numbers and their names from in the output of ceph osd lspools. For example, the first pool that was created corresponds to pool number 1. A fully qualified PG ID has the following form:

{pool-num}.{pg-id}

It typically resembles the following:

1.1701b

To retrieve a list of PGs, run the following command:

ceph pg dump

To format the output in JSON format and save it to a file, run the following command:

ceph pg dump -o {filename} --format=json

To query a specific PG, run the following command:

ceph pg {poolnum}.{pg-id} query

Ceph will output the query in JSON format.

The following subsections describe the most common PG states in detail.

Creating

PGs are created when you create a pool: the command that creates a pool specifies the total number of PGs for that pool, and when the pool is created all of those PGs are created as well. Ceph will echo creating while it is creating PGs. After the PG(s) are created, the OSDs that are part of a PG’s Acting Set will peer. Once peering is complete, the PG status should be active+clean. This status means that Ceph clients begin writing to the PG.

Peering

When a PG peers, the OSDs that store the replicas of its data converge on an agreed state of the data and metadata within that PG. When peering is complete, those OSDs agree about the state of that PG. However, completion of the peering process does NOT mean that each replica has the latest contents.

Authoritative History

Ceph will NOT acknowledge a write operation to a client until that write operation is persisted by every OSD in the Acting Set. 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.

Given an accurate record of each acknowledged write operation, Ceph can construct a new authoritative history of the PG--that is, a complete and fully ordered set of operations that, if performed, would bring an OSD’s copy of the PG up to date.

Active

After Ceph has completed the peering process, a PG should become active. The active state means that the data in the PG is generally available for read and write operations in the primary and replica OSDs.

Clean

When a PG is in the clean state, all OSDs holding its data and metadata have successfully peered and there are no stray replicas. Ceph has replicated all objects in the PG 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 PG 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 that a PG can be active+degraded is that an OSD can be active even if it doesn’t yet hold all of the PG’s objects. If an OSD goes down, Ceph marks each PG assigned to the OSD as degraded. The PGs must peer again when the OSD comes back online. However, a client can still write a new object to a degraded PG if it is active.

If an OSD is down and the degraded condition persists, Ceph might 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 determined by mon_osd_down_out_interval, which is set to 600 seconds by default.

A PG can also be in the degraded state because there are one or more objects that Ceph expects to find in the PG but that Ceph cannot find. Although you cannot read or write to unfound objects, you can still access all of the other objects in the degraded PG.

Recovering

Ceph was designed for fault-tolerance, because hardware and other server problems are expected or even routine. When an OSD goes down, its contents might fall behind the current state of other replicas in the PGs. When the OSD has returned to the up state, the contents of the PGs must be updated to reflect that current state. During that time period, the OSD might be in a recovering state.

Recovery is not always trivial, because a hardware failure might cause a cascading failure of multiple OSDs. For example, a network switch for a rack or cabinet might fail, which can cause the OSDs of a number of host machines to fall behind the current state of the cluster. In such a scenario, general recovery is possible only if each of the OSDs recovers after the fault has been resolved.]

Ceph provides a number of settings that determine how the cluster balances the resource contention between the need to process new service requests and the need to recover data objects and restore the PGs 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_thread_timeout setting determines the duration of a thread timeout, because multiple OSDs might fail, restart, and re-peer at staggered rates. The osd_recovery_max_active setting limits the number of recovery requests an OSD can entertain simultaneously, in order to prevent the OSD from failing to serve. The osd_recovery_max_chunk setting limits the size of the recovered data chunks, in order to prevent network congestion.

Back Filling

When a new OSD joins the cluster, CRUSH will reassign PGs from OSDs that are already in the cluster to the newly added OSD. It can put excessive load on the new OSD to force it to immediately accept the reassigned PGs. Back filling the OSD with the PGs allows this process to begin in the background. After the backfill operations have completed, the new OSD will begin serving requests as soon as it is ready.

During the backfill operations, you might see one of several states: backfill_wait indicates that a backfill operation is pending, but is not yet underway; backfilling indicates that a backfill operation is currently underway; and backfill_toofull indicates that a backfill operation was requested but couldn’t be completed due to insufficient storage capacity. When a PG cannot be backfilled, it might be considered incomplete.

The backfill_toofull state might be transient. It might happen that, as PGs are moved around, space becomes available. The backfill_toofull state is similar to backfill_wait in that backfill operations can proceed as soon as conditions change.

Ceph provides a number of settings to manage the load spike associated with the reassignment of PGs to an OSD (especially a new OSD). The osd_max_backfills setting specifies the maximum number of concurrent backfills to and from an OSD (default: 1; note you cannot change this if the mClock scheduler is active, unless you set osd_mclock_override_recovery_settings = true, see mClock backfill). The backfill_full_ratio setting allows an OSD to refuse a backfill request if the OSD is approaching its full ratio (default: 90%). This setting can be changed with the ceph osd set-backfillfull-ratio command. If an OSD refuses a backfill request, the osd_backfill_retry_interval setting allows an OSD to retry the request after a certain interval (default: 30 seconds). OSDs can also set osd_backfill_scan_min and osd_backfill_scan_max in order to manage scan intervals (default: 64 and 512, respectively).

Remapped

When the Acting Set that services a PG changes, the data migrates from the old Acting Set to the new Acting Set. Because it might take time for the new primary OSD to begin servicing requests, the old primary OSD might be required to continue servicing requests until the PG data migration is complete. After data migration has completed, the mapping uses the primary OSD of the new Acting Set.

Stale

Although Ceph uses heartbeats in order to ensure that hosts and daemons are running, the ceph-osd daemons might enter a stuck state where they are not reporting statistics in a timely manner (for example, there might be a temporary network fault). By default, OSD daemons report their PG, up through, boot, and failure statistics every half second (that is, in accordance with a value of 0.5), which is more frequent than the reports defined by the heartbeat thresholds. If the primary OSD of a PG’s Acting Set fails to report to the monitor or if other OSDs have reported the primary OSD down, the monitors will mark the PG 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 a while, however, seeing PGs in the stale state indicates that the primary OSD for those PGs is down or not reporting PG statistics to the monitor.

Identifying Troubled PGs

As previously noted, a PG is not necessarily having problems just because its state is not active+clean. When PGs are stuck, this might indicate that Ceph cannot perform self-repairs. The stuck states include:

  • Unclean: PGs contain objects that have not been replicated the desired number of times. Under normal conditions, it can be assumed that these PGs are recovering.

  • Inactive: PGs cannot process reads or writes because they are waiting for an OSD that has the most up-to-date data to come back up.

  • Stale: PG are in an unknown state, because the OSDs that host them have not reported to the monitor cluster for a certain period of time (determined by mon_osd_report_timeout).

To identify stuck PGs, run the following command:

ceph pg dump_stuck [unclean|inactive|stale|undersized|degraded]

For more detail, see Placement Group Subsystem. To troubleshoot stuck PGs, see Troubleshooting PG Errors.

Finding an Object Location

To store object data in the Ceph Object Store, a Ceph client must:

  1. Set an object name

  2. Specify a pool

The Ceph client retrieves the latest cluster map, the CRUSH algorithm calculates how to map the object to a PG, and then the algorithm calculates how to dynamically assign the PG to an OSD. To find the object location given only the object name and the pool name, run a command of the following form:

ceph osd map {poolname} {object-name} [namespace]

Exercise: Locate an Object

As an exercise, let’s create an object. We can specify an object name, a path to a test file that contains some object data, and a pool name by 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, run the following command:

rados -p data ls

To identify the object location, run the following commands:

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' (1) object 'test-object-1' -> pg 1.d1743484 (1.4) -> up ([0,1], p0) acting ([0,1], p0)

To remove the test object, simply delete it by running 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 spares you the burden of manually performing the migration. For details, see the Architecture section.