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