=================== Using salt at scale =================== The focus of this tutorial will be building a Salt infrastructure for handling large numbers of minions. This will include tuning, topology, and best practices. For how to install the saltmaster please go here: `Installing saltstack `_ .. note:: This tutorial is intended for large installations, although these same settings won't hurt, it may not be worth the complexity to smaller installations. When used with minions, the term 'many' refers to at least a thousand and 'a few' always means 500. For simplicity reasons, this tutorial will default to the standard ports used by salt. The Master ========== The most common problems on the salt-master are: 1. too many minions authing at once 2. too many minions re-authing at once 3. too many minions re-connecting at once 4. too many minions returning at once 5. too few resources (CPU/HDD) The first three are all "thundering herd" problems. To mitigate these issues we must configure the minions to back-off appropriately when the master is under heavy load. The fourth is caused by masters with little hardware resources in combination with a possible bug in ZeroMQ. At least thats what it looks like till today (`Issue 118651 `_, `Issue 5948 `_, `Mail thread `_) To fully understand each problem, it is important to understand, how salt works. Very briefly, the saltmaster offers two services to the minions. - a job publisher on port 4505 - an open port 4506 to receive the minions returns All minions are always connected to the publisher on port 4505 and only connect to the open return port 4506 if necessary. On an idle master, there will only be connections on port 4505. 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, when all minion keys have already been accepted, the framework is being tested and parameters change frequently in the masters configuration file. In a few cases (master restart, remove minion key, etc.) the salt-master generates a new AES-key to encrypt its publications with. The minions aren't notified of this but will realize this 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 .. code-block:: yaml 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) .. code-block:: yaml recon_default: 100ms recon_max: 5000 recon_randomize: True - recon_default: the default value the socket should use, i.e. 100ms - recon_max: the max value that the socket should use as a delay before trying to reconnect - 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. .. code-block:: yaml 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). .. code-block:: bash 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 .. code-block:: text 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 .. code-block:: bash $ salt * test.ping 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 salts batch mode: .. code-block:: bash $ salt * test.ping -b 50 This will only address 50 minions at once while looping through all addressed minions. Too few resources ================= The masters resources always have to match the environment. There is no way to give good advise without knowing the environment the master is supposed to run in. But here are some general tuning tips for different situations: The master is CPU bound ----------------------- Salt uses RSA-Key-Pairs on the masters and minions end. Both generate 4096 bit key-pairs on first start. While the key-size for the master is currently not configurable, the minions keysize can be configured with different key-sizes. For example with a 2048 bit key: .. code-block:: yaml keysize: 2048 With thousands of decryptions, the amount of time that can be saved on the masters end should not be neglected. See here for reference: `Pull Request 9235 `_ how much influence the key-size can have. Downsizing the salt-masters key is not that important, because the minions do not encrypt as many messages as the master does. 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: .. code-block:: yaml 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 .. code-block:: yaml keep_jobs: 24 .. code-block:: text 250 jobs/day * 2000 minions returns = 500.000 files a day If no job history is needed, the job cache can be disabled: .. code-block:: yaml job_cache: False If the job cache is necessary there are (currently) 2 options: - ext_job_cache: this will have the minions store their return data directly into a returner (not sent through the master) - master_job_cache (New in `2014.7.0`): this will make the master store the job data using a returner (instead of the local job cache on disk).