swift/doc/source/admin_guide.rst

76 KiB

Administrator's Guide

Defining Storage Policies

Defining your Storage Policies is very easy to do with Swift. It is important that the administrator understand the concepts behind Storage Policies before actually creating and using them in order to get the most benefit out of the feature and, more importantly, to avoid having to make unnecessary changes once a set of policies have been deployed to a cluster.

It is highly recommended that the reader fully read and comprehend overview_policies before proceeding with administration of policies. Plan carefully and it is suggested that experimentation be done first on a non-production cluster to be certain that the desired configuration meets the needs of the users. See upgrade-policy before planning the upgrade of your existing deployment.

Following is a high level view of the very few steps it takes to configure policies once you have decided what you want to do:

  1. Define your policies in /etc/swift/swift.conf
  2. Create the corresponding object rings
  3. Communicate the names of the Storage Policies to cluster users

For a specific example that takes you through these steps, please see policies_saio

Managing the Rings

You may build the storage rings on any server with the appropriate version of Swift installed. Once built or changed (rebalanced), you must distribute the rings to all the servers in the cluster. Storage rings contain information about all the Swift storage partitions and how they are distributed between the different nodes and disks.

Swift 1.6.0 is the last version to use a Python pickle format. Subsequent versions use a different serialization format. Rings generated by Swift versions 1.6.0 and earlier may be read by any version, but rings generated after 1.6.0 may only be read by Swift versions greater than 1.6.0. So when upgrading from version 1.6.0 or earlier to a version greater than 1.6.0, either upgrade Swift on your ring building server last after all Swift nodes have been successfully upgraded, or refrain from generating rings until all Swift nodes have been successfully upgraded.

If you need to downgrade from a version of swift greater than 1.6.0 to a version less than or equal to 1.6.0, first downgrade your ring-building server, generate new rings, push them out, then continue with the rest of the downgrade.

For more information see overview_ring.

Removing a device from the ring:

swift-ring-builder <builder-file> remove <ip_address>/<device_name>

Removing a server from the ring:

swift-ring-builder <builder-file> remove <ip_address>

Adding devices to the ring:

See ring-preparing

See what devices for a server are in the ring:

swift-ring-builder <builder-file> search <ip_address>

Once you are done with all changes to the ring, the changes need to be "committed":

swift-ring-builder <builder-file> rebalance

Once the new rings are built, they should be pushed out to all the servers in the cluster.

Optionally, if invoked as 'swift-ring-builder-safe' the directory containing the specified builder file will be locked (via a .lock file in the parent directory). This provides a basic safe guard against multiple instances of the swift-ring-builder (or other utilities that observe this lock) from attempting to write to or read the builder/ring files while operations are in progress. This can be useful in environments where ring management has been automated but the operator still needs to interact with the rings manually.

If the ring builder is not producing the balances that you are expecting, you can gain visibility into what it's doing with the --debug flag.:

swift-ring-builder <builder-file> rebalance --debug

This produces a great deal of output that is mostly useful if you are either (a) attempting to fix the ring builder, or (b) filing a bug against the ring builder.

You may notice in the rebalance output a 'dispersion' number. What this number means is explained in ring_dispersion but in essence is the percentage of partitions in the ring that have too many replicas within a particular failure domain. You can ask 'swift-ring-builder' what the dispersion is with:

swift-ring-builder <builder-file> dispersion

This will give you the percentage again, if you want a detailed view of the dispersion simply add a --verbose:

swift-ring-builder <builder-file> dispersion --verbose

This will not only display the percentage but will also display a dispersion table that lists partition dispersion by tier. You can use this table to figure out were you need to add capacity or to help tune an ring_overload value.

Now let's take an example with 1 region, 3 zones and 4 devices. Each device has the same weight, and the dispersion --verbose might show the following:

Dispersion is 50.000000, Balance is 0.000000, Overload is 0.00%
Required overload is 33.333333%
Worst tier is 50.000000 (r1z3)
--------------------------------------------------------------------------
Tier                           Parts      %    Max     0     1     2     3
--------------------------------------------------------------------------
r1                               256   0.00      3     0     0     0   256
r1z1                             192   0.00      1    64   192     0     0
r1z1-127.0.0.1                   192   0.00      1    64   192     0     0
r1z1-127.0.0.1/sda               192   0.00      1    64   192     0     0
r1z2                             192   0.00      1    64   192     0     0
r1z2-127.0.0.2                   192   0.00      1    64   192     0     0
r1z2-127.0.0.2/sda               192   0.00      1    64   192     0     0
r1z3                             256  50.00      1     0   128   128     0
r1z3-127.0.0.3                   256  50.00      1     0   128   128     0
r1z3-127.0.0.3/sda               192   0.00      1    64   192     0     0
r1z3-127.0.0.3/sdb               192   0.00      1    64   192     0     0

The first line reports that there are 256 partitions with 3 copies in region 1; and this is an expected output in this case (single region with 3 replicas) as reported by the "Max" value.

However, there is some inbalance in the cluster, more precisely in zone 3. The "Max" reports a maximum of 1 copy in this zone; however 50.00% of the partitions are storing 2 replicas in this zone (which is somewhat expected, because there are more disks in this zone).

You can now either add more capacity to the other zones, decrease the total weight in zone 3 or set the overload to a value greater than 33.333333% -only as much overload as needed will be used.

Scripting Ring Creation

You can create scripts to create the account and container rings and rebalance. Here's an example script for the Account ring. Use similar commands to create a make-container-ring.sh script on the proxy server node.

  1. Create a script file called make-account-ring.sh on the proxy server node with the following content:

    #!/bin/bash
    cd /etc/swift
    rm -f account.builder account.ring.gz backups/account.builder backups/account.ring.gz
    swift-ring-builder account.builder create 18 3 1
    swift-ring-builder account.builder add r1z1-<account-server-1>:6202/sdb1 1
    swift-ring-builder account.builder add r1z2-<account-server-2>:6202/sdb1 1
    swift-ring-builder account.builder rebalance

    You need to replace the values of <account-server-1>, <account-server-2>, etc. with the IP addresses of the account servers used in your setup. You can have as many account servers as you need. All account servers are assumed to be listening on port 6202, and have a storage device called "sdb1" (this is a directory name created under /drives when we setup the account server). The "z1", "z2", etc. designate zones, and you can choose whether you put devices in the same or different zones. The "r1" designates the region, with different regions specified as "r1", "r2", etc.

  2. Make the script file executable and run it to create the account ring file:

    chmod +x make-account-ring.sh
    sudo ./make-account-ring.sh
  3. Copy the resulting ring file /etc/swift/account.ring.gz to all the account server nodes in your Swift environment, and put them in the /etc/swift directory on these nodes. Make sure that every time you change the account ring configuration, you copy the resulting ring file to all the account nodes.

Handling System Updates

It is recommended that system updates and reboots are done a zone at a time. This allows the update to happen, and for the Swift cluster to stay available and responsive to requests. It is also advisable when updating a zone, let it run for a while before updating the other zones to make sure the update doesn't have any adverse effects.

Handling Drive Failure

In the event that a drive has failed, the first step is to make sure the drive is unmounted. This will make it easier for swift to work around the failure until it has been resolved. If the drive is going to be replaced immediately, then it is just best to replace the drive, format it, remount it, and let replication fill it up.

After the drive is unmounted, make sure the mount point is owned by root (root:root 755). This ensures that rsync will not try to replicate into the root drive once the failed drive is unmounted.

If the drive can't be replaced immediately, then it is best to leave it unmounted, and set the device weight to 0. This will allow all the replicas that were on that drive to be replicated elsewhere until the drive is replaced. Once the drive is replaced, the device weight can be increased again. Setting the device weight to 0 instead of removing the drive from the ring gives Swift the chance to replicate data from the failing disk too (in case it is still possible to read some of the data).

Setting the device weight to 0 (or removing a failed drive from the ring) has another benefit: all partitions that were stored on the failed drive are distributed over the remaining disks in the cluster, and each disk only needs to store a few new partitions. This is much faster compared to replicating all partitions to a single, new disk. It decreases the time to recover from a degraded number of replicas significantly, and becomes more and more important with bigger disks.

Handling Server Failure

If a server is having hardware issues, it is a good idea to make sure the swift services are not running. This will allow Swift to work around the failure while you troubleshoot.

If the server just needs a reboot, or a small amount of work that should only last a couple of hours, then it is probably best to let Swift work around the failure and get the machine fixed and back online. When the machine comes back online, replication will make sure that anything that is missing during the downtime will get updated.

If the server has more serious issues, then it is probably best to remove all of the server's devices from the ring. Once the server has been repaired and is back online, the server's devices can be added back into the ring. It is important that the devices are reformatted before putting them back into the ring as it is likely to be responsible for a different set of partitions than before.

Detecting Failed Drives

It has been our experience that when a drive is about to fail, error messages will spew into /var/log/kern.log. There is a script called swift-drive-audit that can be run via cron to watch for bad drives. If errors are detected, it will unmount the bad drive, so that Swift can work around it. The script takes a configuration file with the following settings:

[drive-audit]

Option Default Description
log_facility LOG_LOCAL0 Syslog log facility
log_level INFO Log level
device_dir /srv/node Directory devices are mounted under

minutes

60

Number of minutes to look back in /var/log/kern.log

error_limit

1

Number of errors to find before a device is unmounted

log_file_pattern

/var/log/kern*

Location of the log file with globbing pattern to check against device errors

regex_pattern_X

(see below)

Regular expression patterns to be used to locate device blocks with errors in the log file

The default regex pattern used to locate device blocks with errors are berrorb.*b(sd[a-z]{1,2}d?)b and b(sd[a-z]{1,2}d?)b.*berrorb. One is able to overwrite the default above by providing new expressions using the format regex_pattern_X = regex_expression, where X is a number.

This script has been tested on Ubuntu 10.04 and Ubuntu 12.04, so if you are using a different distro or OS, some care should be taken before using in production.

Preventing Disk Full Scenarios

Prevent disk full scenarios by ensuring that the proxy-server blocks PUT requests and rsync prevents replication to the specific drives.

You can prevent proxy-server PUT requests to low space disks by ensuring fallocate_reserve is set in the object-server.conf. By default, fallocate_reserve is set to 1%. This blocks PUT requests that leave the free disk space below 1% of the disk.

In order to prevent rsync replication to specific drives, firstly setup rsync_module per disk in your object-replicator. Set this in object-server.conf:

[object-replicator]
rsync_module = {replication_ip}::object_{device}

Set the individual drives in rsync.conf. For example:

[object_sda]
max connections = 4
lock file = /var/lock/object_sda.lock

[object_sdb]
max connections = 4
lock file = /var/lock/object_sdb.lock

Finally, monitor the disk space of each disk and adjust the rsync max connections per drive to -1. We recommend utilising your existing monitoring solution to achieve this. The following is an example script:

#!/usr/bin/env python
import os
import errno

RESERVE = 500 * 2 ** 20  # 500 MiB

DEVICES = '/srv/node1'

path_template = '/etc/rsync.d/disable_%s.conf'
config_template = '''
[object_%s]
max connections = -1
'''

def disable_rsync(device):
    with open(path_template % device, 'w') as f:
        f.write(config_template.lstrip() % device)


def enable_rsync(device):
    try:
        os.unlink(path_template % device)
    except OSError as e:
        # ignore file does not exist
        if e.errno != errno.ENOENT:
            raise


for device in os.listdir(DEVICES):
    path = os.path.join(DEVICES, device)
    st = os.statvfs(path)
    free = st.f_bavail * st.f_frsize
    if free < RESERVE:
        disable_rsync(device)
    else:
        enable_rsync(device)

For the above script to work, ensure /etc/rsync.d/ conf files are included, by specifying &include in your rsync.conf file:

&include /etc/rsync.d

Use this in conjunction with a cron job to periodically run the script, for example:

# /etc/cron.d/devicecheck
* * * * * root /some/path/to/disable_rsync.py

Dispersion Report

There is a swift-dispersion-report tool for measuring overall cluster health. This is accomplished by checking if a set of deliberately distributed containers and objects are currently in their proper places within the cluster.

For instance, a common deployment has three replicas of each object. The health of that object can be measured by checking if each replica is in its proper place. If only 2 of the 3 is in place the object's heath can be said to be at 66.66%, where 100% would be perfect.

A single object's health, especially an older object, usually reflects the health of that entire partition the object is in. If we make enough objects on a distinct percentage of the partitions in the cluster, we can get a pretty valid estimate of the overall cluster health. In practice, about 1% partition coverage seems to balance well between accuracy and the amount of time it takes to gather results.

The first thing that needs to be done to provide this health value is create a new account solely for this usage. Next, we need to place the containers and objects throughout the system so that they are on distinct partitions. The swift-dispersion-populate tool does this by making up random container and object names until they fall on distinct partitions. Last, and repeatedly for the life of the cluster, we need to run the swift-dispersion-report tool to check the health of each of these containers and objects.

These tools need direct access to the entire cluster and to the ring files (installing them on a proxy server will probably do). Both swift-dispersion-populate and swift-dispersion-report use the same configuration file, /etc/swift/dispersion.conf. Example conf file:

[dispersion]
auth_url = http://localhost:8080/auth/v1.0
auth_user = test:tester
auth_key = testing
endpoint_type = internalURL

There are also options for the conf file for specifying the dispersion coverage (defaults to 1%), retries, concurrency, etc. though usually the defaults are fine. If you want to use keystone v3 for authentication there are options like auth_version, user_domain_name, project_domain_name and project_name.

Once the configuration is in place, run swift-dispersion-populate to populate the containers and objects throughout the cluster.

Now that those containers and objects are in place, you can run swift-dispersion-report to get a dispersion report, or the overall health of the cluster. Here is an example of a cluster in perfect health:

$ swift-dispersion-report
Queried 2621 containers for dispersion reporting, 19s, 0 retries
100.00% of container copies found (7863 of 7863)
Sample represents 1.00% of the container partition space

Queried 2619 objects for dispersion reporting, 7s, 0 retries
100.00% of object copies found (7857 of 7857)
Sample represents 1.00% of the object partition space

Now I'll deliberately double the weight of a device in the object ring (with replication turned off) and rerun the dispersion report to show what impact that has:

$ swift-ring-builder object.builder set_weight d0 200
$ swift-ring-builder object.builder rebalance
...
$ swift-dispersion-report
Queried 2621 containers for dispersion reporting, 8s, 0 retries
100.00% of container copies found (7863 of 7863)
Sample represents 1.00% of the container partition space

Queried 2619 objects for dispersion reporting, 7s, 0 retries
There were 1763 partitions missing one copy.
77.56% of object copies found (6094 of 7857)
Sample represents 1.00% of the object partition space

You can see the health of the objects in the cluster has gone down significantly. Of course, I only have four devices in this test environment, in a production environment with many many devices the impact of one device change is much less. Next, I'll run the replicators to get everything put back into place and then rerun the dispersion report:

... start object replicators and monitor logs until they're caught up ...
$ swift-dispersion-report
Queried 2621 containers for dispersion reporting, 17s, 0 retries
100.00% of container copies found (7863 of 7863)
Sample represents 1.00% of the container partition space

Queried 2619 objects for dispersion reporting, 7s, 0 retries
100.00% of object copies found (7857 of 7857)
Sample represents 1.00% of the object partition space

You can also run the report for only containers or objects:

$ swift-dispersion-report --container-only
Queried 2621 containers for dispersion reporting, 17s, 0 retries
100.00% of container copies found (7863 of 7863)
Sample represents 1.00% of the container partition space

$ swift-dispersion-report --object-only
Queried 2619 objects for dispersion reporting, 7s, 0 retries
100.00% of object copies found (7857 of 7857)
Sample represents 1.00% of the object partition space

Alternatively, the dispersion report can also be output in json format. This allows it to be more easily consumed by third party utilities:

$ swift-dispersion-report -j
{"object": {"retries:": 0, "missing_two": 0, "copies_found": 7863, "missing_one": 0, "copies_expected": 7863, "pct_found": 100.0, "overlapping": 0, "missing_all": 0}, "container": {"retries:": 0, "missing_two": 0, "copies_found": 12534, "missing_one": 0, "copies_expected": 12534, "pct_found": 100.0, "overlapping": 15, "missing_all": 0}}

Note that you may select which storage policy to use by setting the option '--policy-name silver' or '-P silver' (silver is the example policy name here). If no policy is specified, the default will be used per the swift.conf file. When you specify a policy the containers created also include the policy index, thus even when running a container_only report, you will need to specify the policy not using the default.

Geographically Distributed Clusters

Swift's default configuration is currently designed to work in a single region, where a region is defined as a group of machines with high-bandwidth, low-latency links between them. However, configuration options exist that make running a performant multi-region Swift cluster possible.

For the rest of this section, we will assume a two-region Swift cluster: region 1 in San Francisco (SF), and region 2 in New York (NY). Each region shall contain within it 3 zones, numbered 1, 2, and 3, for a total of 6 zones.

read_affinity

This setting, combined with sorting_method setting, makes the proxy server prefer local backend servers for GET and HEAD requests over non-local ones. For example, it is preferable for an SF proxy server to service object GET requests by talking to SF object servers, as the client will receive lower latency and higher throughput.

By default, Swift randomly chooses one of the three replicas to give to the client, thereby spreading the load evenly. In the case of a geographically-distributed cluster, the administrator is likely to prioritize keeping traffic local over even distribution of results. This is where the read_affinity setting comes in.

Example:

[app:proxy-server]
sorting_method = affinity
read_affinity = r1=100

This will make the proxy attempt to service GET and HEAD requests from backends in region 1 before contacting any backends in region 2. However, if no region 1 backends are available (due to replica placement, failed hardware, or other reasons), then the proxy will fall back to backend servers in other regions.

Example:

[app:proxy-server]
sorting_method = affinity
read_affinity = r1z1=100, r1=200

This will make the proxy attempt to service GET and HEAD requests from backends in region 1 zone 1, then backends in region 1, then any other backends. If a proxy is physically close to a particular zone or zones, this can provide bandwidth savings. For example, if a zone corresponds to servers in a particular rack, and the proxy server is in that same rack, then setting read_affinity to prefer reads from within the rack will result in less traffic between the top-of-rack switches.

The read_affinity setting may contain any number of region/zone specifiers; the priority number (after the equals sign) determines the ordering in which backend servers will be contacted. A lower number means higher priority.

Note that read_affinity only affects the ordering of primary nodes (see ring docs for definition of primary node), not the ordering of handoff nodes.

write_affinity and write_affinity_node_count

This setting makes the proxy server prefer local backend servers for object PUT requests over non-local ones. For example, it may be preferable for an SF proxy server to service object PUT requests by talking to SF object servers, as the client will receive lower latency and higher throughput. However, if this setting is used, note that a NY proxy server handling a GET request for an object that was PUT using write affinity may have to fetch it across the WAN link, as the object won't immediately have any replicas in NY. However, replication will move the object's replicas to their proper homes in both SF and NY.

Note that only object PUT requests are affected by the write_affinity setting; POST, GET, HEAD, DELETE, OPTIONS, and account/container PUT requests are not affected.

This setting lets you trade data distribution for throughput. If write_affinity is enabled, then object replicas will initially be stored all within a particular region or zone, thereby decreasing the quality of the data distribution, but the replicas will be distributed over fast WAN links, giving higher throughput to clients. Note that the replicators will eventually move objects to their proper, well-distributed homes.

The write_affinity setting is useful only when you don't typically read objects immediately after writing them. For example, consider a workload of mainly backups: if you have a bunch of machines in NY that periodically write backups to Swift, then odds are that you don't then immediately read those backups in SF. If your workload doesn't look like that, then you probably shouldn't use write_affinity.

The write_affinity_node_count setting is only useful in conjunction with write_affinity; it governs how many local object servers will be tried before falling back to non-local ones.

Example:

[app:proxy-server]
write_affinity = r1
write_affinity_node_count = 2 * replicas

Assuming 3 replicas, this configuration will make object PUTs try storing the object's replicas on up to 6 disks ("2 * replicas") in region 1 ("r1"). Proxy server tries to find 3 devices for storing the object. While a device is unavailable, it queries the ring for the 4th device and so on until 6th device. If the 6th disk is still unavailable, the last replica will be sent to other region. It doesn't mean there'll have 6 replicas in region 1.

You should be aware that, if you have data coming into SF faster than your link to NY can transfer it, then your cluster's data distribution will get worse and worse over time as objects pile up in SF. If this happens, it is recommended to disable write_affinity and simply let object PUTs traverse the WAN link, as that will naturally limit the object growth rate to what your WAN link can handle.

Cluster Telemetry and Monitoring

Various metrics and telemetry can be obtained from the account, container, and object servers using the recon server middleware and the swift-recon cli. To do so update your account, container, or object servers pipelines to include recon and add the associated filter config.

object-server.conf sample:

[pipeline:main]
pipeline = recon object-server

[filter:recon]
use = egg:swift#recon
recon_cache_path = /var/cache/swift

container-server.conf sample:

[pipeline:main]
pipeline = recon container-server

[filter:recon]
use = egg:swift#recon
recon_cache_path = /var/cache/swift

account-server.conf sample:

[pipeline:main]
pipeline = recon account-server

[filter:recon]
use = egg:swift#recon
recon_cache_path = /var/cache/swift

The recon_cache_path simply sets the directory where stats for a few items will be stored. Depending on the method of deployment you may need to create this directory manually and ensure that swift has read/write access.

Finally, if you also wish to track asynchronous pending on your object servers you will need to setup a cronjob to run the swift-recon-cron script periodically on your object servers:

*/5 * * * * swift /usr/bin/swift-recon-cron /etc/swift/object-server.conf

Once the recon middleware is enabled, a GET request for "/recon/<metric>" to the backend object server will return a JSON-formatted response:

fhines@ubuntu:~$ curl -i http://localhost:6030/recon/async
HTTP/1.1 200 OK
Content-Type: application/json
Content-Length: 20
Date: Tue, 18 Oct 2011 21:03:01 GMT

{"async_pending": 0}

Note that the default port for the object server is 6200, except on a Swift All-In-One installation, which uses 6010, 6020, 6030, and 6040.

The following metrics and telemetry are currently exposed:

Request URI Description
/recon/load returns 1,5, and 15 minute load average
/recon/mem returns /proc/meminfo
/recon/mounted returns ALL currently mounted filesystems
/recon/unmounted returns all unmounted drives if mount_check = True
/recon/diskusage returns disk utilization for storage devices
/recon/ringmd5 returns object/container/account ring md5sums
/recon/quarantined returns # of quarantined objects/accounts/containers
/recon/sockstat returns consumable info from /proc/net/sockstat|6
/recon/devices returns list of devices and devices dir i.e. /srv/node
/recon/async returns count of async pending
/recon/replication returns object replication info (for backward compatibility)
/recon/replication/<type> returns replication info for given type (account, container, object)
/recon/auditor/<type> returns auditor stats on last reported scan for given type (account, container, object)
/recon/updater/<type> returns last updater sweep times for given type (container, object)

Note that 'object_replication_last' and 'object_replication_time' in object replication info are considered to be transitional and will be removed in the subsequent releases. Use 'replication_last' and 'replication_time' instead.

This information can also be queried via the swift-recon command line utility:

fhines@ubuntu:~$ swift-recon -h
Usage:
        usage: swift-recon <server_type> [-v] [--suppress] [-a] [-r] [-u] [-d]
        [-l] [-T] [--md5] [--auditor] [--updater] [--expirer] [--sockstat]

        <server_type>   account|container|object
        Defaults to object server.

        ex: swift-recon container -l --auditor


Options:
  -h, --help            show this help message and exit
  -v, --verbose         Print verbose info
  --suppress            Suppress most connection related errors
  -a, --async           Get async stats
  -r, --replication     Get replication stats
  --auditor             Get auditor stats
  --updater             Get updater stats
  --expirer             Get expirer stats
  -u, --unmounted       Check cluster for unmounted devices
  -d, --diskusage       Get disk usage stats
  -l, --loadstats       Get cluster load average stats
  -q, --quarantined     Get cluster quarantine stats
  --md5                 Get md5sum of servers ring and compare to local copy
  --sockstat            Get cluster socket usage stats
  -T, --time            Check time synchronization
  --all                 Perform all checks. Equal to
                        -arudlqT --md5 --sockstat --auditor --updater
                        --expirer --driveaudit --validate-servers
  -z ZONE, --zone=ZONE  Only query servers in specified zone
  -t SECONDS, --timeout=SECONDS
                        Time to wait for a response from a server
  --swiftdir=SWIFTDIR   Default = /etc/swift

For example, to obtain container replication info from all hosts in zone "3":

fhines@ubuntu:~$ swift-recon container -r --zone 3
===============================================================================
--> Starting reconnaissance on 1 hosts
===============================================================================
[2012-04-02 02:45:48] Checking on replication
[failure] low: 0.000, high: 0.000, avg: 0.000, reported: 1
[success] low: 486.000, high: 486.000, avg: 486.000, reported: 1
[replication_time] low: 20.853, high: 20.853, avg: 20.853, reported: 1
[attempted] low: 243.000, high: 243.000, avg: 243.000, reported: 1

Reporting Metrics to StatsD

If you have a StatsD server running, Swift may be configured to send it real-time operational metrics. To enable this, set the following configuration entries (see the sample configuration files):

log_statsd_host = localhost
log_statsd_port = 8125
log_statsd_default_sample_rate = 1.0
log_statsd_sample_rate_factor = 1.0
log_statsd_metric_prefix =                [empty-string]

If log_statsd_host is not set, this feature is disabled. The default values for the other settings are given above. The log_statsd_host can be a hostname, an IPv4 address, or an IPv6 address (not surrounded with brackets, as this is unnecessary since the port is specified separately). If a hostname resolves to an IPv4 address, an IPv4 socket will be used to send StatsD UDP packets, even if the hostname would also resolve to an IPv6 address.

The sample rate is a real number between 0 and 1 which defines the probability of sending a sample for any given event or timing measurement. This sample rate is sent with each sample to StatsD and used to multiply the value. For example, with a sample rate of 0.5, StatsD will multiply that counter's value by 2 when flushing the metric to an upstream monitoring system (Graphite, Ganglia, etc.).

Some relatively high-frequency metrics have a default sample rate less than one. If you want to override the default sample rate for all metrics whose default sample rate is not specified in the Swift source, you may set log_statsd_default_sample_rate to a value less than one. This is NOT recommended (see next paragraph). A better way to reduce StatsD load is to adjust log_statsd_sample_rate_factor to a value less than one. The log_statsd_sample_rate_factor is multiplied to any sample rate (either the global default or one specified by the actual metric logging call in the Swift source) prior to handling. In other words, this one tunable can lower the frequency of all StatsD logging by a proportional amount.

To get the best data, start with the default log_statsd_default_sample_rate and log_statsd_sample_rate_factor values of 1 and only lower log_statsd_sample_rate_factor if needed. The log_statsd_default_sample_rate should not be used and remains for backward compatibility only.

The metric prefix will be prepended to every metric sent to the StatsD server For example, with:

log_statsd_metric_prefix = proxy01

the metric proxy-server.errors would be sent to StatsD as proxy01.proxy-server.errors. This is useful for differentiating different servers when sending statistics to a central StatsD server. If you run a local StatsD server per node, you could configure a per-node metrics prefix there and leave log_statsd_metric_prefix blank.

Note that metrics reported to StatsD are counters or timing data (which are sent in units of milliseconds). StatsD usually expands timing data out to min, max, avg, count, and 90th percentile per timing metric, but the details of this behavior will depend on the configuration of your StatsD server. Some important "gauge" metrics may still need to be collected using another method. For example, the object-server.async_pendings StatsD metric counts the generation of async_pendings in real-time, but will not tell you the current number of async_pending container updates on disk at any point in time.

Note also that the set of metrics collected, their names, and their semantics are not locked down and will change over time.

Metrics for `account-auditor`:

Metric Name Description

account-auditor.errors

Count of audit runs (across all account databases) which caught an Exception.

account-auditor.passes Count of individual account databases which passed audit.
account-auditor.failures Count of individual account databases which failed audit.
account-auditor.timing Timing data for individual account database audits.

Metrics for `account-reaper`:

Metric Name Description
account-reaper.errors Count of devices failing the mount check.
account-reaper.timing Timing data for each reap_account() call.

account-reaper.return_codes.X

Count of HTTP return codes from various operations (e.g. object listing, container deletion, etc.). The value for X is the first digit of the return code (2 for 201, 4 for 404, etc.).

account-reaper.containers_failures Count of failures to delete a container.
account-reaper.containers_deleted Count of containers successfully deleted.

account-reaper.containers_remaining

Count of containers which failed to delete with zero successes.

account-reaper.containers_possibly_remaining

Count of containers which failed to delete with at least one success.

account-reaper.objects_failures Count of failures to delete an object.
account-reaper.objects_deleted Count of objects successfully deleted.

account-reaper.objects_remaining

Count of objects which failed to delete with zero successes.

account-reaper.objects_possibly_remaining

Count of objects which failed to delete with at least one success.

Metrics for account-server ("Not Found" is not considered an error and requests which increment errors are not included in the timing data):

Metric Name Description

account-server.DELETE.errors.timing

Timing data for each DELETE request resulting in an error: bad request, not mounted, missing timestamp.

account-server.DELETE.timing

Timing data for each DELETE request not resulting in an error.

account-server.PUT.errors.timing

Timing data for each PUT request resulting in an error: bad request, not mounted, conflict, recently-deleted.

account-server.PUT.timing

Timing data for each PUT request not resulting in an error.

account-server.HEAD.errors.timing

Timing data for each HEAD request resulting in an error: bad request, not mounted.

account-server.HEAD.timing

Timing data for each HEAD request not resulting in an error.

account-server.GET.errors.timing

Timing data for each GET request resulting in an error: bad request, not mounted, bad delimiter, account listing limit too high, bad accept header.

account-server.GET.timing

Timing data for each GET request not resulting in an error.

account-server.REPLICATE.errors.timing

Timing data for each REPLICATE request resulting in an error: bad request, not mounted.

account-server.REPLICATE.timing

Timing data for each REPLICATE request not resulting in an error.

account-server.POST.errors.timing

Timing data for each POST request resulting in an error: bad request, bad or missing timestamp, not mounted.

account-server.POST.timing

Timing data for each POST request not resulting in an error.

Metrics for `account-replicator`:

Metric Name Description
account-replicator.diffs Count of syncs handled by sending differing rows.

account-replicator.diff_caps

Count of "diffs" operations which failed because "max_diffs" was hit.

account-replicator.no_changes Count of accounts found to be in sync.

account-replicator.hashmatches

Count of accounts found to be in sync via hash comparison (broker.merge_syncs was called).

account-replicator.rsyncs

Count of completely missing accounts which were sent via rsync.

account-replicator.remote_merges

Count of syncs handled by sending entire database via rsync.

account-replicator.attempts Count of database replication attempts.

account-replicator.failures

Count of database replication attempts which failed due to corruption (quarantined) or inability to read as well as attempts to individual nodes which failed.

account-replicator.removes.<device>

Count of databases on <device> deleted because the delete_timestamp was greater than the put_timestamp and the database had no rows or because it was successfully sync'ed to other locations and doesn't belong here anymore.

account-replicator.successes

Count of replication attempts to an individual node which were successful.

account-replicator.timing

Timing data for each database replication attempt not resulting in a failure.

Metrics for `container-auditor`:

Metric Name Description

container-auditor.errors

Incremented when an Exception is caught in an audit pass (only once per pass, max).

container-auditor.passes Count of individual containers passing an audit.
container-auditor.failures Count of individual containers failing an audit.
container-auditor.timing Timing data for each container audit.

Metrics for `container-replicator`:

Metric Name Description
container-replicator.diffs Count of syncs handled by sending differing rows.

container-replicator.diff_caps

Count of "diffs" operations which failed because "max_diffs" was hit.

container-replicator.no_changes Count of containers found to be in sync.

container-replicator.hashmatches

Count of containers found to be in sync via hash comparison (broker.merge_syncs was called).

container-replicator.rsyncs

Count of completely missing containers where were sent via rsync.

container-replicator.remote_merges

Count of syncs handled by sending entire database via rsync.

container-replicator.attempts Count of database replication attempts.

container-replicator.failures

Count of database replication attempts which failed due to corruption (quarantined) or inability to read as well as attempts to individual nodes which failed.

container-replicator.removes.<device>

Count of databases deleted on <device> because the delete_timestamp was greater than the put_timestamp and the database had no rows or because it was successfully sync'ed to other locations and doesn't belong here anymore.

container-replicator.successes

Count of replication attempts to an individual node which were successful.

container-replicator.timing

Timing data for each database replication attempt not resulting in a failure.

Metrics for container-server ("Not Found" is not considered an error and requests which increment errors are not included in the timing data):

Metric Name Description

container-server.DELETE.errors.timing

Timing data for DELETE request errors: bad request, not mounted, missing timestamp, conflict.

container-server.DELETE.timing

Timing data for each DELETE request not resulting in an error.

container-server.PUT.errors.timing

Timing data for PUT request errors: bad request, missing timestamp, not mounted, conflict.

container-server.PUT.timing

Timing data for each PUT request not resulting in an error.

container-server.HEAD.errors.timing

Timing data for HEAD request errors: bad request, not mounted.

container-server.HEAD.timing

Timing data for each HEAD request not resulting in an error.

container-server.GET.errors.timing

Timing data for GET request errors: bad request, not mounted, parameters not utf8, bad accept header.

container-server.GET.timing

Timing data for each GET request not resulting in an error.

container-server.REPLICATE.errors.timing

Timing data for REPLICATE request errors: bad request, not mounted.

container-server.REPLICATE.timing

Timing data for each REPLICATE request not resulting in an error.

container-server.POST.errors.timing

Timing data for POST request errors: bad request, bad x-container-sync-to, not mounted.

container-server.POST.timing

Timing data for each POST request not resulting in an error.

Metrics for `container-sync`:

Metric Name Description

container-sync.skips

Count of containers skipped because they don't have sync'ing enabled.

container-sync.failures Count of failures sync'ing of individual containers.
container-sync.syncs Count of individual containers sync'ed successfully.

container-sync.deletes

Count of container database rows sync'ed by deletion.

container-sync.deletes.timing

Timing data for each container database row synchronization via deletion.

container-sync.puts Count of container database rows sync'ed by PUTing.

container-sync.puts.timing

Timing data for each container database row synchronization via PUTing.

Metrics for `container-updater`:

Metric Name Description

container-updater.successes

Count of containers which successfully updated their account.

container-updater.failures

Count of containers which failed to update their account.

container-updater.no_changes

Count of containers which didn't need to update their account.

container-updater.timing

Timing data for processing a container; only includes timing for containers which needed to update their accounts (i.e. "successes" and "failures" but not "no_changes").

Metrics for `object-auditor`:

Metric Name Description
object-auditor.quarantines Count of objects failing audit and quarantined.
object-auditor.errors Count of errors encountered while auditing objects.

object-auditor.timing

Timing data for each object audit (does not include any rate-limiting sleep time for max_files_per_second, but does include rate-limiting sleep time for max_bytes_per_second).

Metrics for `object-expirer`:

Metric Name Description
object-expirer.objects Count of objects expired.

object-expirer.errors

Count of errors encountered while attempting to expire an object.

object-expirer.timing

Timing data for each object expiration attempt, including ones resulting in an error.

Metrics for `object-reconstructor`:

Metric Name Description

object-reconstructor.partition.delete.count.<device>

A count of partitions on <device> which were reconstructed and synced to another node because they didn't belong on this node. This metric is tracked per-device to allow for "quiescence detection" for object reconstruction activity on each device.

object-reconstructor.partition.delete.timing

Timing data for partitions reconstructed and synced to another node because they didn't belong on this node. This metric is not tracked per device.

object-reconstructor.partition.update.count.<device>

A count of partitions on <device> which were reconstructed and synced to another node, but also belong on this node. As with delete.count, this metric is tracked per-device.

object-reconstructor.partition.update.timing

Timing data for partitions reconstructed which also belong on this node. This metric is not tracked per-device.

object-reconstructor.suffix.hashes

Count of suffix directories whose hash (of filenames) was recalculated.

object-reconstructor.suffix.syncs Count of suffix directories reconstructed with ssync.

Metrics for `object-replicator`:

Metric Name Description

object-replicator.partition.delete.count.<device>

A count of partitions on <device> which were replicated to another node because they didn't belong on this node. This metric is tracked per-device to allow for "quiescence detection" for object replication activity on each device.

object-replicator.partition.delete.timing

Timing data for partitions replicated to another node because they didn't belong on this node. This metric is not tracked per device.

object-replicator.partition.update.count.<device>

A count of partitions on <device> which were replicated to another node, but also belong on this node. As with delete.count, this metric is tracked per-device.

object-replicator.partition.update.timing

Timing data for partitions replicated which also belong on this node. This metric is not tracked per-device.

object-replicator.suffix.hashes

Count of suffix directories whose hash (of filenames) was recalculated.

object-replicator.suffix.syncs Count of suffix directories replicated with rsync.

Metrics for `object-server`:

Metric Name Description

object-server.quarantines

Count of objects (files) found bad and moved to quarantine.

object-server.async_pendings

Count of container updates saved as async_pendings (may result from PUT or DELETE requests).

object-server.POST.errors.timing

Timing data for POST request errors: bad request, missing timestamp, delete-at in past, not mounted.

object-server.POST.timing

Timing data for each POST request not resulting in an error.

object-server.PUT.errors.timing

Timing data for PUT request errors: bad request, not mounted, missing timestamp, object creation constraint violation, delete-at in past.

object-server.PUT.timeouts Count of object PUTs which exceeded max_upload_time.

object-server.PUT.timing

Timing data for each PUT request not resulting in an error.

object-server.PUT.<device>.timing

Timing data per kB transferred (ms/kB) for each non-zero-byte PUT request on each device. Monitoring problematic devices, higher is bad.

object-server.GET.errors.timing

Timing data for GET request errors: bad request, not mounted, header timestamps before the epoch, precondition failed. File errors resulting in a quarantine are not counted here.

object-server.GET.timing

Timing data for each GET request not resulting in an error. Includes requests which couldn't find the object (including disk errors resulting in file quarantine).

object-server.HEAD.errors.timing

Timing data for HEAD request errors: bad request, not mounted.

object-server.HEAD.timing

Timing data for each HEAD request not resulting in an error. Includes requests which couldn't find the object (including disk errors resulting in file quarantine).

object-server.DELETE.errors.timing

Timing data for DELETE request errors: bad request, missing timestamp, not mounted, precondition failed. Includes requests which couldn't find or match the object.

object-server.DELETE.timing

Timing data for each DELETE request not resulting in an error.

object-server.REPLICATE.errors.timing

Timing data for REPLICATE request errors: bad request, not mounted.

object-server.REPLICATE.timing

Timing data for each REPLICATE request not resulting in an error.

Metrics for `object-updater`:

Metric Name Description

object-updater.errors

Count of drives not mounted or async_pending files with an unexpected name.

object-updater.timing

Timing data for object sweeps to flush async_pending container updates. Does not include object sweeps which did not find an existing async_pending storage directory.

object-updater.quarantines

Count of async_pending container updates which were corrupted and moved to quarantine.

object-updater.successes Count of successful container updates.
object-updater.failures Count of failed container updates.

object-updater.unlinks

Count of async_pending files unlinked. An async_pending file is unlinked either when it is successfully processed or when the replicator sees that there is a newer async_pending file for the same object.

Metrics for proxy-server (in the table, <type> is the proxy-server controller responsible for the request and will be one of "account", "container", or "object"):

Metric Name Description

proxy-server.errors

Count of errors encountered while serving requests before the controller type is determined. Includes invalid Content-Length, errors finding the internal controller to handle the request, invalid utf8, and bad URLs.

proxy-server.<type>.handoff_count

Count of node hand-offs; only tracked if log_handoffs is set in the proxy-server config.

proxy-server.<type>.handoff_all_count

Count of times only hand-off locations were utilized; only tracked if log_handoffs is set in the proxy-server config.

proxy-server.<type>.client_timeouts

Count of client timeouts (client did not read within client_timeout seconds during a GET or did not supply data within client_timeout seconds during a PUT).

proxy-server.<type>.client_disconnects

Count of detected client disconnects during PUT operations (does NOT include caught Exceptions in the proxy-server which caused a client disconnect).

Metrics for proxy-logging middleware (in the table, <type> is either the proxy-server controller responsible for the request: "account", "container", "object", or the string "SOS" if the request came from the Swift Origin Server middleware. The <verb> portion will be one of "GET", "HEAD", "POST", "PUT", "DELETE", "COPY", "OPTIONS", or "BAD_METHOD". The list of valid HTTP methods is configurable via the log_statsd_valid_http_methods config variable and the default setting yields the above behavior):

Metric Name Description

proxy-server.<type>.<verb>.<status>.timing

Timing data for requests, start to finish. The <status> portion is the numeric HTTP status code for the request (e.g. "200" or "404").

proxy-server.<type>.GET.<status>.first-byte.timing

Timing data up to completion of sending the response headers (only for GET requests). <status> and <type> are as for the main timing metric.

proxy-server.<type>.<verb>.<status>.xfer

This counter metric is the sum of bytes transferred in (from clients) and out (to clients) for requests. The <type>, <verb>, and <status> portions of the metric are just like the main timing metric.

The proxy-logging middleware also groups these metrics by policy. The <policy-index> portion represents a policy index):

Metric Name Description

proxy-server.object.policy.<policy-index>.<verb>.<status>.timing

Timing data for requests, aggregated by policy index.

proxy-server.object.policy.<policy-index>.GET.<status>.first-byte.timing

Timing data up to completion of sending the response headers, aggregated by policy index.

proxy-server.object.policy.<policy-index>.<verb>.<status>.xfer

Sum of bytes transferred in and out, aggregated by policy index.

Metrics for tempauth middleware (in the table, <reseller_prefix> represents the actual configured reseller_prefix or "NONE" if the reseller_prefix is the empty string):

Metric Name Description

tempauth.<reseller_prefix>.unauthorized

Count of regular requests which were denied with HTTPUnauthorized.

tempauth.<reseller_prefix>.forbidden

Count of regular requests which were denied with HTTPForbidden.

tempauth.<reseller_prefix>.token_denied Count of token requests which were denied.
tempauth.<reseller_prefix>.errors Count of errors.

Debugging Tips and Tools

When a request is made to Swift, it is given a unique transaction id. This id should be in every log line that has to do with that request. This can be useful when looking at all the services that are hit by a single request.

If you need to know where a specific account, container or object is in the cluster, swift-get-nodes will show the location where each replica should be.

If you are looking at an object on the server and need more info, swift-object-info will display the account, container, replica locations and metadata of the object.

If you are looking at a container on the server and need more info, swift-container-info will display all the information like the account, container, replica locations and metadata of the container.

If you are looking at an account on the server and need more info, swift-account-info will display the account, replica locations and metadata of the account.

If you want to audit the data for an account, swift-account-audit can be used to crawl the account, checking that all containers and objects can be found.

Managing Services

Swift services are generally managed with swift-init. the general usage is swift-init <service> <command>, where service is the swift service to manage (for example object, container, account, proxy) and command is one of:

Command Description
start Start the service
stop Stop the service
restart Restart the service
shutdown Attempt to gracefully shutdown the service
reload Attempt to gracefully restart the service

A graceful shutdown or reload will finish any current requests before completely stopping the old service. There is also a special case of swift-init all <command>, which will run the command for all swift services.

In cases where there are multiple configs for a service, a specific config can be managed with swift-init <service>.<config> <command>. For example, when a separate replication network is used, there might be /etc/swift/object-server/public.conf for the object server and /etc/swift/object-server/replication.conf for the replication services. In this case, the replication services could be restarted with swift-init object-server.replication restart.

Object Auditor

On system failures, the XFS file system can sometimes truncate files it's trying to write and produce zero-byte files. The object-auditor will catch these problems but in the case of a system crash it would be advisable to run an extra, less rate limited sweep to check for these specific files. You can run this command as follows: swift-object-auditor /path/to/object-server/config/file.conf once -z 1000 "-z" means to only check for zero-byte files at 1000 files per second.

At times it is useful to be able to run the object auditor on a specific device or set of devices. You can run the object-auditor as follows: swift-object-auditor /path/to/object-server/config/file.conf once --devices=sda,sdb

This will run the object auditor on only the sda and sdb devices. This param accepts a comma separated list of values.

Object Replicator

At times it is useful to be able to run the object replicator on a specific device or partition. You can run the object-replicator as follows: swift-object-replicator /path/to/object-server/config/file.conf once --devices=sda,sdb

This will run the object replicator on only the sda and sdb devices. You can likewise run that command with --partitions. Both params accept a comma separated list of values. If both are specified they will be ANDed together. These can only be run in "once" mode.

Swift Orphans

Swift Orphans are processes left over after a reload of a Swift server.

For example, when upgrading a proxy server you would probably finish with a swift-init proxy-server reload or /etc/init.d/swift-proxy reload. This kills the parent proxy server process and leaves the child processes running to finish processing whatever requests they might be handling at the time. It then starts up a new parent proxy server process and its children to handle new incoming requests. This allows zero-downtime upgrades with no impact to existing requests.

The orphaned child processes may take a while to exit, depending on the length of the requests they were handling. However, sometimes an old process can be hung up due to some bug or hardware issue. In these cases, these orphaned processes will hang around forever. swift-orphans can be used to find and kill these orphans.

swift-orphans with no arguments will just list the orphans it finds that were started more than 24 hours ago. You shouldn't really check for orphans until 24 hours after you perform a reload, as some requests can take a long time to process. swift-orphans -k TERM will send the SIG_TERM signal to the orphans processes, or you can kill -TERM the pids yourself if you prefer.

You can run swift-orphans --help for more options.

Swift Oldies

Swift Oldies are processes that have just been around for a long time. There's nothing necessarily wrong with this, but it might indicate a hung process if you regularly upgrade and reload/restart services. You might have so many servers that you don't notice when a reload/restart fails; swift-oldies can help with this.

For example, if you upgraded and reloaded/restarted everything 2 days ago, and you've already cleaned up any orphans with swift-orphans, you can run swift-oldies -a 48 to find any Swift processes still around that were started more than 2 days ago and then investigate them accordingly.

Custom Log Handlers

Swift supports setting up custom log handlers for services by specifying a comma-separated list of functions to invoke when logging is setup. It does so via the log_custom_handlers configuration option. Logger hooks invoked are passed the same arguments as Swift's get_logger function (as well as the getLogger and LogAdapter object):

Name Description
conf Configuration dict to read settings from
name Name of the logger received
log_to_console (optional) Write log messages to console on stderr
log_route Route for the logging received
fmt Override log format received
logger The logging.getLogger object
adapted_logger The LogAdapter object

A basic example that sets up a custom logger might look like the following:

def my_logger(conf, name, log_to_console, log_route, fmt, logger,
              adapted_logger):
    my_conf_opt = conf.get('some_custom_setting')
    my_handler = third_party_logstore_handler(my_conf_opt)
    logger.addHandler(my_handler)

See custom-logger-hooks-label for sample use cases.

Securing OpenStack Swift

Please refer to the security guides at: