Using Salt at scale¶

Too many minions authing¶

When the Minion service is first started up, it will connect to its Master's publisher on port 4505. If too many minions are started at once, this can cause a "thundering herd". This can be avoided by not starting too many minions at once.

The connection itself usually isn't the culprit, the more likely cause of master-side issues is the authentication that the Minion must do with the Master. If the Master is too heavily loaded to handle the auth request it will time it out. The Minion will then wait acceptance_wait_time to retry. If acceptance_wait_time_max is set then the Minion will increase its wait time by the acceptance_wait_time each subsequent retry until reaching acceptance_wait_time_max.

Too many minions re-authing¶

This is most likely to happen in the testing phase of a Salt deployment, when all Minion keys have already been accepted, but the framework is being tested and parameters are frequently changed in the Salt Master's configuration file(s).

The Salt Master generates a new AES key to encrypt its publications at certain events such as a Master restart or the removal of a Minion key. If you are encountering this problem of too many minions re-authing against the Master, you will need to recalibrate your setup to reduce the rate of events like a Master restart or Minion key removal (salt-key -d).

When the Master generates a new AES key, the minions aren't notified of this but will discover it on the next pub job they receive. When the Minion receives such a job it will then re-auth with the Master. Since Salt does minion-side filtering this means that all the minions will re-auth on the next command published on the master-- causing another "thundering herd". This can be avoided by setting the

random_reauth_delay: 60

in the minions configuration file to a higher value and stagger the amount of re-auth attempts. Increasing this value will of course increase the time it takes until all minions are reachable via Salt commands.

Too many minions re-connecting¶

By default the zmq socket will re-connect every 100ms which for some larger installations may be too quick. This will control how quickly the TCP session is re-established, but has no bearing on the auth load.

To tune the minions sockets reconnect attempts, there are a few values in the sample configuration file (default values)

recon_default: 1000
recon_max: 5000
recon_randomize: True

• recon_default: the default value the socket should use, i.e. 1000. This value is in milliseconds. (1000ms = 1 second)

• recon_max: the max value that the socket should use as a delay before trying to reconnect This value is in milliseconds. (5000ms = 5 seconds)

• recon_randomize: enables randomization between recon_default and recon_max

To tune this values to an existing environment, a few decision have to be made.

1. How long can one wait, before the minions should be online and reachable via Salt?

2. How many reconnects can the Master handle without a syn flood?

These questions can not be answered generally. Their answers depend on the hardware and the administrators requirements.

Here is an example scenario with the goal, to have all minions reconnect within a 60 second time-frame on a Salt Master service restart.

recon_default: 1000
recon_max: 59000
recon_randomize: True

Each Minion will have a randomized reconnect value between 'recon_default' and 'recon_default + recon_max', which in this example means between 1000ms and 60000ms (or between 1 and 60 seconds). The generated random-value will be doubled after each attempt to reconnect (ZeroMQ default behavior).

Lets say the generated random value is 11 seconds (or 11000ms).

reconnect 1: wait 11 seconds
reconnect 2: wait 22 seconds
reconnect 3: wait 33 seconds
reconnect 4: wait 44 seconds
reconnect 5: wait 55 seconds
reconnect 6: wait time is bigger than 60 seconds (recon_default + recon_max)
reconnect 7: wait 11 seconds
reconnect 8: wait 22 seconds
reconnect 9: wait 33 seconds
reconnect x: etc.

With a thousand minions this will mean

1000/60 = ~16

round about 16 connection attempts a second. These values should be altered to values that match your environment. Keep in mind though, that it may grow over time and that more minions might raise the problem again.

Too many minions returning at once¶

This can also happen during the testing phase, if all minions are addressed at once with

$ salt * disk.usage

it may cause thousands of minions trying to return their data to the Salt Master open port 4506. Also causing a flood of syn-flood if the Master can't handle that many returns at once.

This can be easily avoided with Salt's batch mode:

$ salt * disk.usage -b 50

This will only address 50 minions at once while looping through all addressed minions.

The Master is CPU bound¶

In installations with large or with complex pillar files, it is possible for the master to exhibit poor performance as a result of having to render many pillar files at once. This exhibit itself in a number of ways, both as high load on the master and on minions which block on waiting for their pillar to be delivered to them.

To reduce pillar rendering times, it is possible to cache pillars on the master. To do this, see the set of master configuration options which are prefixed with pillar_cache.

If many pillars are encrypted using gpg renderer, it is possible to cache GPG data. To do this, see the set of master configuration options which are prefixed with gpg_cache.

The Master is disk IO bound¶

By default, the Master saves every Minion's return for every job in its job-cache. The cache can then be used later, to lookup results for previous jobs. The default directory for this is:

cachedir: /var/cache/salt

and then in the /proc directory.

Each job return for every Minion is saved in a single file. Over time this directory can grow quite large, depending on the number of published jobs. The amount of files and directories will scale with the number of jobs published and the retention time defined by

keep_jobs_seconds: 86400

250 jobs/day * 2000 minions returns = 500,000 files a day

job_cache: False