88
**In short the rule is:** "repositories must hold revisions' parent
89
inventories and their new texts (or else all texts for those revisions)."
91
This is sometimes called "the stacking invariant."
96
A stacked repository needs to be capable of generating a complete stream
97
for the revisions it does hold without access to its fallback
98
repositories [#]_. "Complete" here means that the stream for a revision (or
99
set of revisions) can be inserted into a repository that already contains
100
the parent(s) of that revision, and that repository will have a fully
101
usable copy of that revision: a working tree can be built for that
104
Assuming for a moment the stream has the necessary inventory, signature
105
and CHK records to have a usable revision, what texts are required to have
106
a usable revision? The simple way to satisfy the requirement is to have
107
*every* text for every revision at the stacking boundary. Thus the
108
revisions at the stacking boundary and all their descendants have their
109
texts present and so can be fully reconstructed. But this is expensive:
110
it implies each stacked repository much contain *O(tree)* data even for a
111
single revision of a 1-line change, and also implies transferring
112
*O(tree)* data to fetch that revision.
114
Because the goal is a usable revision *when added to a repository with the
115
parent revision(s)* most of those texts will be redundant. The minimal
116
set that is needed is just those texts that are new in the revisions in
117
our repository. However, we need enough inventory data to be able to
118
determine that set of texts. So to make this possible every revision must
119
have its parent inventories present so that the inventory delta between
120
revisions can be calculated, and of course the CHK pages associated with
121
that delta. In fact the entire inventory does not need to be present,
122
just enough of it to find the delta (assuming a repository format, like
123
2a, that allows only part of an inventory to be stored). Thus the stacked
124
repository can contain only *O(changes)* data [#]_ and still deliver
125
complete streams of that data.
127
What about revisions at the stacking boundary with more than one parent?
128
All of their parent revisions must be present, as a client may ask for a
129
stream up to any parent, not just the left-hand parent. If any parent is
130
absent then all texts must be present instead. Otherwise there will be
131
the strange situation where some fetches of a revision will succeed and
132
others fail depending the precise details of the fetch.
134
Implications for fetching
135
-------------------------
137
Fetches must retrieve the records necessary to satisfy that rule. The
138
stream source will attempt to send the necessary records, and the stream
139
sink will check for any missing records and make a second fetch for just
140
those missing records before committing the write group.
142
Our repository implementations check this constraint is satisfied before
143
committing a write group, to prevent a bad stream from creating a corrupt
144
repository. So a fetch from a bad source (e.g. a damaged repository, or a
145
buggy foreign-format import) may trigger ``BzrCheckError`` during
146
``commit_write_group``.
148
To fetch from a stacked repository via a smart server, the smart client:
150
* first fetches a stream of as many of the requested revisions as possible
151
from the initial repository,
152
* then while there are still missing revisions and untried fallback
153
repositories fetches the outstanding revisions from the next fallback
154
until either all revisions have been found (success) or the list of
155
fallbacks has been exhausted (failure).
158
.. [#] This is not just a theoretical concern. The smart server always
159
opens repositories without opening fallbacks, as it cannot assume it
160
can access the fallbacks that the client can.
162
.. [#] Actually *O(changes)* isn't quite right in practice. In the
163
current implementation the fulltext of a changed file must be
164
transferred, not just a delta, so a 1-line change to a 10MB file will
165
still transfer 10MB of text data. This is because current formats
166
require records' compression parents to be present in the same
86
171
vim: ft=rst tw=74 ai