github.com/makyo/juju@v0.0.0-20160425123129-2608902037e9/doc/death-and-destruction.txt (about) 1 Death and Destruction 2 ===================== 3 4 This document describes in detail the operations associated with the destruction 5 and removal of the fundamental state entities, and what agents are responsible 6 for those operations. 7 8 Each entity has an associated destroy-* command. The precise implications of 9 destruction differ by entity, but there are common features: 10 11 * Only Alive entities can be destroyed; if destruction is already in progress, 12 as evidenced by an entity not being Alive, its "destruction" is a no-op. 13 * Entities might be removed immediately when they are destroyed, but this is not 14 guaranteed. 15 * If an entity is not removed immediately when it is destroyed, its eventual 16 removal is very likely; but it is not currently guaranteed, for the 17 following reasons: 18 * Hardware failure, even when detected and corrected by a Provisioner, can 19 lead to unremovable relations, because the redeployed unit doesn't know 20 what relations it's in. This would be fixable by making the unit agent 21 always leave the scope of relations when they're detected; or, probably 22 better, by using actual remote scope membership state to track relation 23 membership (rather than using the existence of a local directory, whose 24 true intent is to track the membership of *other* units, as a proxy). 25 This is actually a pretty serious BUG and should be addressed soon; 26 neither proposed solution is very challenging. 27 * Undetected hardware failure is annoying, and can block progress at any 28 time, but can be observed via additional monitoring and resolved via out- 29 of-band termination of borked resources, which should be sufficient to 30 get the system moving again (assuming the above bug is fixed). 31 * Unknown problems in juju, in which agents fail to fulfil the duties laid 32 out in this document, could block progress at any time. Assuming a 33 version of the agent code which does not exhibit the problem exists, it 34 should always be possible to work around this situation by upgrading the 35 agent; and, if that fails, by terminating the underlying provider 36 resources out-of-band, as above, and waiting for the new agent version 37 to be deployed on a fresh system (with the same caveat as above). 38 * In light of the preceding two points, we don't *have* to implement 39 "--force" options for `juju destroy-machine` and `juju destroy-unit`. 40 This is good, because it will be tricky to implement them well. 41 42 In general, the user can just forget about entities once she's destroyed them; 43 the only caveat is that she may not create new services with the same name, or 44 new relations identical to the destroyed ones, until those entities have 45 finally been removed. 46 47 In rough order of complexity, here's what happens when each entity kind is 48 destroyed. Note that in general the appropriate action is contingent on 49 mutable remote state, and many operations must be expressed as a transaction 50 involving several documents: the state API must be prepared to handle aborted 51 transactions and either diagnose definite failure or retry until the operation 52 succeeds (or, perhaps, finally error out pleading excessive contention). 53 54 55 juju destroy-machine 56 -------------------- 57 58 Destroying a machine involves a single transaction defined as follows: 59 60 * If the machine is not Alive, abort without error. 61 * If the machine is the last one with JobManageModel, or has any assigned 62 units, abort with an appropriate error. 63 * Set the machine to Dying. 64 65 When a machine becomes Dying, the following operation occurs: 66 67 * The machine's agent sets the machine to Dead. 68 69 When a machine becomes Dead, the following operations occur: 70 71 * The machine's agent terminates itself and refuses to run again. 72 * A Provisioner (a task running in some other machine agent) observes the 73 death, decommissions the machine's resources, and removes the machine. 74 75 Removing a machine involves a single transaction defined as follows: 76 77 * If the machine is not Dead, abort with an appropriate error. 78 * Delete the machine document. 79 80 81 juju destroy-unit 82 ----------------- 83 84 Destroying a unit involves a single transaction defined as follows: 85 86 * If the unit is not Alive, abort without error. 87 * Set the unit to Dying. 88 89 When a unit becomes Dying, the following operations occur: 90 91 * The unit's agent leaves the scopes of all its relations. Note that this is 92 a potentially complex sequence of operations and may take some time; in 93 particular, any hooks that fail while the unit is leaving relations and 94 stopping the charm will suspend this sequence until resolved (just like 95 when the unit is Alive). 96 * The unit's agent then sets the unit to Dead. 97 98 When a unit becomes Dead, the following operations occur: 99 100 * The unit's agent terminates itself and refuses to run again. 101 * The agent of the entity that deployed the unit (that is: a machine agent, 102 for a principal unit; or, for a subordinate unit, the agent of a principal 103 unit) observes the death, recalls the unit, and removes it. 104 105 Removing a unit involves a single transaction, defined as follows: 106 107 * If the unit is a principal unit, unassign the unit from its machine. 108 * If the unit is a subordinate unit, unassign it from its principal unit. 109 * Delete the unit document. 110 * If its service is Alive, or has at least two units, or is in at least 111 one relation, decrement the service's unit count; otherwise remove the 112 service. 113 114 115 juju destroy-relation 116 --------------------- 117 118 Destroying a relation involves a single transaction defined as follows: 119 120 * If the relation is not Alive, abort without error. 121 * If any unit is in scope, set the relation to Dying. 122 * Otherwise: 123 * If the relation destruction is a direct user request, decrement the 124 relation counts of both services. 125 * If the relation destruction is an immediate consequence of service 126 destruction, decrement the reference count of the counterpart service 127 alone. (This is because the service destruction logic is responsible 128 for the relation count of the service being destroyed.) 129 * Delete the relation document. 130 * Mark the relation's unit settings documents for future cleanup. 131 * This is done by creating a single document for the attention of 132 some other part of the system (BUG: which doesn't exist), that is 133 then responsible for mass-deleting the (potentially large number 134 of) settings documents. This completely bypasses the mgo/txn 135 mechanism, but we don't care because those documents are guaranteed 136 to be unreferenced and unwatched, by virtue of the relation's prior 137 removal. 138 139 When a relation is set to Dying, the following operations occur: 140 141 * Every unit agent whose unit has entered the scope of that relation 142 observes the change and causes its unit to leave scope. 143 * If the relation has container scope, and no other container-scoped relation 144 between its services is Alive, the unit agents of the subordinate units in 145 the relation will observe the change and destroy their units. 146 147 The Dying relation's document is finally removed in the same transaction in 148 which the last unit leaves its scope. Because this situation involves the 149 relation already being Dying, its services may also be Dying, and so the 150 operations involved are subtly different to those taken above (when we know 151 for sure that the relation -- and hence both services -- are still Alive). 152 153 * Here, "the service" refers to the service of the unit departing scope, and 154 "the counterpart service" refers to the other service in the relation. 155 * Decrement the relation count of the unit's service (we know that service 156 is not ready to be removed, because its unit is responsible for this 157 transaction and the service clearly therefore has a unit count greater 158 than zero). 159 * Delete the relation document. 160 * Mark the relation's unit settings documents for future cleanup. 161 * If the counterpart service (the one that is not the unit's service) is 162 Alive, or has at least one unit, or is in at least two relations, decrement 163 its relation count; otherwise remove the counterpart service. 164 165 166 juju destroy-service 167 -------------------- 168 169 Destroying a service involves a single transaction defined as follows: 170 171 * If the service is not Alive, abort without error. 172 * If the service is in any relations, do the following for each one: 173 * If the relation is already Dying, skip it. 174 * If the relation is Alive, destroy the relation without modifying the 175 service's relation count. If the relation's destruction implies its 176 removal, increment a local removed-relations counter instead. 177 * If the service's unit count is greater than 0, or if the value of the 178 aforementioned removal counter is less than the service's relation count, 179 we know that some entity will still hold a reference to the service after 180 the transaction completes, so we set the service to Dying and decrement 181 its relation count by the value of the removal counter. 182 * Otherwise, remove the service immediately, because we know that no 183 reference to the service will survive the transaction. 184 185 When a service becomes Dying, the following operations occur: 186 187 * Every unit agent of the service observes the change and destroys its unit. 188 189 The Dying service's document is finally removed in the same transaction that 190 removes the last entity referencing that service. This could be either the 191 removal of the last unit in the service, or the removal of the last relation 192 the service is in, as described above. To remove a service, the following 193 operations must occur in a single transaction: 194 195 * Remove the service document. 196 * Remove the service's settings document.