| Age | Commit message (Collapse) | Author |
|---|
|
If we create the gen document before saving the actual document in
couch, we may run into problems if more than one client is syncing and
trying to save documents with the same id at the same time.
By moving the gen document creation to after the actual document save in
couch, we rely on couch/u1db resolution of conflicts before actually
allocating a new generation, and the problem above doesn't occur.
|
|
|
|
test_processing_order aims to check that unordered docs wont be
processed, but if we let the pool start and advance Twisted LoopingCall
clock right before calling the processing method manually, the process
method will run concurrently and cause a race condition issue.
|
|
|
|
|
|
"tox -e pep8" runs it standalone and "tox" includes the pep8 env.
|
|
|
|
The use of a lock to allocate the next generation of a change in couch
backend suffers from at least 2 problems:
1. all modification to the couch database would have to be made through
a soledad server entrypoint, otherwise the lock would have no effect.
2. introducing a lock makes code uglier, harder to debug, and prone to
undesired blocks.
The solution implemented by this commit is not so elegant, but works for
what we need right now. Now, concurrent threads updating the couch
database will race for the allocation of a new generation, and retry
when they fail to do so.
There's no high risk of getting blocked for too much time in the while
loop because (1) there's always one thread that wins (what makes the
expected number of retries to be N/2 if N is the number of concurrent
threads), and (2) the number of concurrent attempts to update the user
database is limited by the number of devices syncing at the same time.
|
|
The previous solution would make use of concurrent get's to couch
backend in a pool of threads to implement the get_docs() and
get_all_docs() CouchDatabase backend methods.
This commit replaces those by a simpler implementation use the
`_all_docs` couchdb view api. It passes all needed IDs to the view and
r etrieves all documents with content in the same request.
A comparison between both implementations shows an improvement of at
least 15 times for large number of documents. The table below shows the
time for different implementations of get_all_docs() for different
number of documents and threads versus _all_docs implementation:
+-------+-----------------+------------------+-------------+
| | threads | _all_docs | improvement |
+-------+-----------------+------------------+-------------+
| 10 | 0.0728030204773 | 0.00782012939453 | 9.3 |
| 100 | 0.609349966049 | 0.0377721786499 | 16.1 |
| 1000 | 5.86522197723 | 0.370730876923 | 15.8 |
| 10000 | 66.1713931561 | 3.61764383316 | 18.3 |
+-------+-----------------+------------------+-------------+
|
|
|
|
|
|
Recent versions of pip will ignore that option and use a cache anyway.
|
|
|
|
|
|
`pytest_addoption` was declared twice making the second declaration
replace the first, thus removing couch url parameter.
|
|
|
|
|
|
|
|
container
|
|
|
|
|
|
|
|
|
|
|
|
Currently the perf tests use pytest-twisted plugin, and this has some
implications in the old tests adapted from u1db that now use trial
classes. Because of that, we exclude perf tests from usual tox calls,
but you can still run them by explicitelly calling `tox perf`.
|
|
|
|
|
|
|
|
perf tests
|
|
|
|
|
|
The couch backend makes use of attachments and multipart structure for
writing the document to the couch database. For that to work, the order
in which attachments are described must match the actual order in which
attachments are written to the couch http stream.
This was not being properly taken care of, and eventually the json
serializer was arbitrarilly ordering the attachments description in a
way that it didn't match the actual order of attachments writing.
This commit fixes that by using json.dumps() sort_keys parameter and
making sure conflicts are always written before content.
|
|
|
|
Design documents are slow and we already have alternatives to all uses
we used to make of them, so this commit completelly removes all usage of
design documents.
|
|
|
|
When compared to plain couch document get, the use of the simplest view
functions takes around double the time, while the use of the simplest
list function can take more than 8 times:
get 100 docs:
total: 0.440337 secs
mean: 0.004403
query 100 views:
total: 0.911425 secs
mean: 0.009114
query 100 lists:
total: 3.711537 secs
mean: 0.037115
Besides that, the current implementation of sync metadata storage over
couch is dependent of timestamps of document puts, what can lead to
metadata corruption if the clock of the system is changed for any
reason.
Because of these reasons, we seek to change the implementation of
database metadata. This commit implements the storage of transaction log
data on couch documents with special ids, in the form "gen-xxxxxxxxxx",
where the x's are replaced by the generation index.
Each generation document holds a dictionary containing the generation,
doc_id and transaction_id for the changed document. For each modified
document, a generation document is inserted holding the transaction
metadata.
|
|
|
|
|
|
|
|
|
|
modifying original PR [0] by cristoph to account for the recent
vendoring of l2db code, which means we no longer depend on u1db/dirspec.
I expect the whole mess about the venv setup to be further simplified
pretty soon, since we are going to merge most of the leap.* packages
into a couple of repos.
[0] https://github.com/leapcode/soledad/pull/327
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
SQLCipher database access errors can raise Soledad exceptions. Database access
and multithreading resources are allocated in different places, so we have to
be careful to close all multithreading mechanismis in case of database access
errors. If we don't, zombie threads may haunt the reactor.
This commit adds SQLCipher exception trapping and Soledad exception raising
for database access errors, while properly shutting down multithreading
resources.
|
|
|
