Mailbox Distribution

Prior to Cyrus IMAP version 2.5.0, when creating a mailbox, should no target partition have been specified, the mailbox is either created on:

  • the configured defaultserver, or

  • the server that is found to have the most free disk space, and

  • the configured defaultpartition, or

  • the most fitting partition if defaultpartition is not configured.

New configuration options are available since Cyrus IMAP 2.5.0, that allow more weighted and better balanced backend and partition selection.

Partition Selection Mode

Among the partitions, how the most fitting one is selected depends on the configured selection mode, using the partition_select_mode setting in imapd.conf(5).

Available Selection Modes

  • random

    Choice is (pseudo-)random. Each partition has the same probability of being selected.

  • freespace-most

    The partition which has the most absolute free space (counted in KiB units) is selected.

    In the example scenario, part2 would be selected as most fitting, since 600 GB of free space is the biggest of all partitions.

  • freespace-percent-most

    The partition which has the most relative free space (counted in percentiles) is selected.

    In the example scenario, part4 would be selected as most fitting, since 70% of free space is the biggest of all partitions.

  • freespace-percent-weighted

    For each partition, the percentage of free space is its weight. Then a weighted choice is performed to select one of those partitions.

    As such, the more free space the partition has, the higher its chances to be selected.

    In the example scenario, the weight of each partition would be:

    • 40 for part1

    • 60 for part2

    • 30 for part3

    • 70 for part4

    The sum of all weights being 200, the probability for each partition to be selected as most fitting is:

    • 20% for part1

    • 30% for part2

    • 15% for part3

    • 35% for part4

    Out of 20 (hypothetically empty) mailboxes to be created, chances are:

    • 4 are created on part1

    • 6 are created on part2

    • 3 are created on part3

    • 7 are created on part4

    Note

    In freespace-percent-weighted mode, partitions percentage usages converge towards 100%. So if they have different usages, those differences will stay and only really diminish upon reaching 100% of usage.

    You may also observe growing differences between partitions usage when those partitions do not have the same total disk space.

freespace-percent-weighted-delta

As for freespace-percent-weighted, a weight is associated to each partition. It is computed as follows:

\[free - leastfree + 0.5\]

Then a weighted choice is performed to select one of those partitions.

As such, considering the percentages of usage, the more the partition is lagging behind the most used partition (which is the one with the lowest percentage of free space), the higher its chances are to be selected.

Note

The added 0.5 in partitions weight causes the selection to get smoother the more partitions get close to each other.

In the example scenario, the weight of each partition would be:

  • For part1: \(40 - 30 + 0.5 = 10.5\)

  • For part2: \(60 - 30 + 0.5 = 30.5\)

  • For part3: \(30 - 30 + 0.5 = 00.5\)

  • For part4: \(70 - 30 + 0.5 = 40.5\)

The sum of all weights amounting to 82, the probability for each partition to be selected as most fitting would be:

  • For part1: 12.8%

  • For part2: 37.2%

  • For part3: 00.6%

  • For part4: 49.4%

Note

In freespace-percent-weighted-delta mode, partitions percentage usages converge towards the most used one, after which partitions usages grow equally.

Special Cases

Q: What happens when two partitions are equal as most fitting?

Suppose you are using the freespace-most selection mode, that two (or more) partitions have the same free size, and that this freespace happens to be the biggest one of all configured partitions.

In that case, only one of those partitions will be selected. You may not know in advance which one will be: it depends of the order in which configured partitions are stored in memory (hashtable). In particular, it may not be the first one that appears listed in your imapd.conf(5) configuration file.

Also note that since the selected partition will now have less free space, it shall not be seen as most fitting next time.

Q: What happens when two partitions point to the same device?

Suppose you are using the freespace-most or freespace-percent-most selection mode, and that two (or more) partitions actually point to a different directory on the same filesystem.

In that case, only one of those partitions will be checked, as if the others were not configured. Cyrus IMAP uses the device id of the partition hierarchy for this. You may not know in advance which one will be: it depends of the order in which configured partitions are stored in memory (hashtable). In particular, it may not be the first one that appears listed in your imapd.conf(5) configuration file.

Excluding Partitions

partition_select_exclude

Partitions listed in the partition_select_exclude setting in imapd.conf(5) are permanently excluded from being selected.

Listed partition's names are separated by spaces and/or commas. Only configured partition names (and not paths) are expected.

With the following configuration in imapd.conf(5), parta and partb would be permanently excluded from being selected:

partition-part1: /path/to/part1
partition-part2: /path/to/part2
partition-part3: /path/to/part3
partition-part4: /path/to/part4
partition-parta: /path/to/parta
partition-partb: /path/to/partb

partition_mode_exclude: parta partb

partition_select_soft_usage_limit

When using a selection mode other than random, partitions are automatically excluded if their usage percentage is beyond the partition_select_soft_usage_limit integer setting.

If all partitions are beyond the configured value, this feature is automatically deactivated. A partition is thus selected as if the setting was not configured.

In the example scenario, setting partition_mode_soft_usage_limit to 50 would exclude partitions part1 and part3 since their disk usages are 60% and 70% respectively.

But setting the option to 20 would have no immediate effect on the selection of a partition, since the usage of all partitions is beyond 20%.

Important

Exclusion is not absolute

Partitions are only excluded when creating a new user mailbox according to the configured selection mode.

If you explicitly specify an excluded partition, the mailbox will be created on that partition.

Mailboxes previously created on such partitions are of course still accessible, and subfolders are by default still created on the same partition as the parent folder.

Partition's Usage Information Reset

By default, partition's usage information is retrieved only once -- when the service first initializes.

If you tend to use the same service instance for a long lapse of time and performs a large amount of mailbox creations, it may be useful to configure the partition_select_usage_reinit setting so that each partition's usage information is refreshed after the configured number of creation requests.

Note

This only concerns selection modes other than random`.

Backend Selection Mode (Cyrus IMAP Murder)

Upon creating a user mailbox, if the target server is not given as extra parameter, the mailbox is created on either;

  • the configured defaultserver backend

  • the most fitting backend (partition) among the servers listed in the serverlist setting, if defaultserver is not configured.

Related Settings

  • serverlist_select_mode

  • serverlist_select_usage_reinit

  • serverlist_select_soft_usage_limit

Available Selection Modes on Frontend

random

The selection is (pseudo-)random. Each backend has the same probability of being selected.

freespace-most

The backend which has the most absolute free space (counted in KiB units) is selected.

Note

The considered free space is the sum of the free space on all available partitions on the backend.

In the example scenario, be1 would be selected as most fitting, since 1000 GB of free space is the most of all backends.

freespace-percent-most

On each backend, only the partition with the most percentage of free space is considered. The selected backend is the one whose partition has the highest percentage of free space.

In the example scenario, be3 would be selected as most fitting, since it has a partition with 80% of free space which is the highest of all backends.

freespace-percent-weighted

On each backend, only the partition with the most percentage of free space is considered: this is the backend weight. Then a weighted choice is performed to select one of the backends.

In the example scenario, the weight of each backend would be:

  • 50 for be1

  • 70 for be2

  • 80 for be3

The sum of all weights being 200, the probability for each backend to be selected as most fitting would be:

  • 25% for be1

  • 35% for be2

  • 40% for be3

freespace-percent-weighted-delta

On each backend, only the partition with the most percentage of free space is considered.

Like with freespace-percent-weighted, a weight is associated to each backend. It is computed as follows:

\[free - leastfree + 0.5\]

Then a weighted choice is performed to select one of the backends.

In the example scenario, the weight of each backend would be:

  • For be1: \(50 - 50 + 0.5 = 0.5\)

  • For be2: \(70 - 50 + 0.5 = 20.5\)

  • For be3: \(80 - 50 + 0.5 = 30.5\)

Then the probability for each backend to be selected as most fitting would be:

  • 1.0% for be1

  • 39.8% for be2

  • 59.2% for be3

Excluding Backends

When using a selection mode other than random, backends are automatically excluded if their considered usage percentage is beyond the serverlist_select_soft_usage_limit integer setting.

Backend's Usage Information Reset

By default backends usage data are retrieved only once upon service initialization.

Note

This only concerns selection modes other than random.

If you tend to use the same service instance for a long lapse of time and performs a large amounts of mailbox creations, it may be useful to configure the serverlist_select_usage_reinit so that the backend's disk usage information is refreshed after the configured number of creation requests.

Back to Features