util: teach lrucachedict to enforce a max total cost Now that lrucachedict entries can have a numeric cost associated with them and we can easily pop the oldest item in the cache, it now becomes relatively trivial to implement support for enforcing a high water mark on the total cost of items in the cache. This commit teaches lrucachedict instances to have a max cost associated with them. When items are inserted, we pop old items until enough "cost" frees up to make room for the new item. This feature is close to zero cost when not used (modulo the insertion regressed introduced by the previous commit): $ ./hg perflrucachedict --size 4 --gets 1000000 --sets 1000000 --mixed 1000000 ! gets ! wall 0.607444 comb 0.610000 user 0.610000 sys 0.000000 (best of 17) ! wall 0.601653 comb 0.600000 user 0.600000 sys 0.000000 (best of 17) ! inserts ! wall 0.678261 comb 0.680000 user 0.680000 sys 0.000000 (best of 14) ! wall 0.685042 comb 0.680000 user 0.680000 sys 0.000000 (best of 15) ! sets ! wall 0.808770 comb 0.800000 user 0.800000 sys 0.000000 (best of 13) ! wall 0.834241 comb 0.830000 user 0.830000 sys 0.000000 (best of 12) ! mixed ! wall 0.782441 comb 0.780000 user 0.780000 sys 0.000000 (best of 13) ! wall 0.803804 comb 0.800000 user 0.800000 sys 0.000000 (best of 13) $ hg perflrucachedict --size 1000 --gets 1000000 --sets 1000000 --mixed 1000000 ! init ! wall 0.006952 comb 0.010000 user 0.010000 sys 0.000000 (best of 418) ! gets ! wall 0.613350 comb 0.610000 user 0.610000 sys 0.000000 (best of 17) ! wall 0.617415 comb 0.620000 user 0.620000 sys 0.000000 (best of 17) ! inserts ! wall 0.701270 comb 0.700000 user 0.700000 sys 0.000000 (best of 15) ! wall 0.700516 comb 0.700000 user 0.700000 sys 0.000000 (best of 15) ! sets ! wall 0.825720 comb 0.830000 user 0.830000 sys 0.000000 (best of 13) ! wall 0.837946 comb 0.840000 user 0.830000 sys 0.010000 (best of 12) ! mixed ! wall 0.821644 comb 0.820000 user 0.820000 sys 0.000000 (best of 13) ! wall 0.850559 comb 0.850000 user 0.850000 sys 0.000000 (best of 12) I reckon the slight slowdown on insert is due to added if checks. For caches with total cost limiting enabled: $ hg perflrucachedict --size 4 --gets 1000000 --sets 1000000 --mixed 1000000 --costlimit 100 ! gets w/ cost limit ! wall 0.598737 comb 0.590000 user 0.590000 sys 0.000000 (best of 17) ! inserts w/ cost limit ! wall 1.694282 comb 1.700000 user 1.700000 sys 0.000000 (best of 6) ! mixed w/ cost limit ! wall 1.157655 comb 1.150000 user 1.150000 sys 0.000000 (best of 9) $ hg perflrucachedict --size 1000 --gets 1000000 --sets 1000000 --mixed 1000000 --costlimit 10000 ! gets w/ cost limit ! wall 0.598526 comb 0.600000 user 0.600000 sys 0.000000 (best of 17) ! inserts w/ cost limit ! wall 37.838315 comb 37.840000 user 37.840000 sys 0.000000 (best of 3) ! mixed w/ cost limit ! wall 18.060198 comb 18.060000 user 18.060000 sys 0.000000 (best of 3) $ hg perflrucachedict --size 1000 --gets 1000000 --sets 1000000 --mixed 1000000 --costlimit 10000 --mixedgetfreq 90 ! gets w/ cost limit ! wall 0.600024 comb 0.600000 user 0.600000 sys 0.000000 (best of 17) ! inserts w/ cost limit ! wall 37.154547 comb 37.120000 user 37.120000 sys 0.000000 (best of 3) ! mixed w/ cost limit ! wall 4.381602 comb 4.380000 user 4.370000 sys 0.010000 (best of 3) The functions we're benchmarking are slightly different, which could move numbers by a few milliseconds. But the slowdown on insert is too great to be explained by that. The slowness is due to insert heavy operations needing to call popoldest() repeatedly when the cache is at capacity. The next commit will address this. Differential Revision: https://phab.mercurial-scm.org/D4503

util: allow lrucachedict to track cost of entries Currently, lrucachedict allows tracking of arbitrary items with the only limit being the total number of items in the cache. Caches can be a lot more useful when they are bound by the size of the items in them rather than the number of elements in the cache. In preparation for teaching lrucachedict to enforce a max size of cached items, we teach lrucachedict to optionally associate a numeric cost value with each node. We purposefully let the caller define their own cost for nodes. This does introduce some overhead. Most of it comes from __setitem__, since that function now calls into insert(), thus introducing Python function call overhead. $ hg perflrucachedict --size 4 --gets 1000000 --sets 1000000 --mixed 1000000 ! gets ! wall 0.599552 comb 0.600000 user 0.600000 sys 0.000000 (best of 17) ! wall 0.614643 comb 0.610000 user 0.610000 sys 0.000000 (best of 17) ! inserts ! <not available> ! wall 0.655817 comb 0.650000 user 0.650000 sys 0.000000 (best of 16) ! sets ! wall 0.540448 comb 0.540000 user 0.540000 sys 0.000000 (best of 18) ! wall 0.805644 comb 0.810000 user 0.810000 sys 0.000000 (best of 13) ! mixed ! wall 0.651556 comb 0.660000 user 0.660000 sys 0.000000 (best of 15) ! wall 0.781357 comb 0.780000 user 0.780000 sys 0.000000 (best of 13) $ hg perflrucachedict --size 1000 --gets 1000000 --sets 1000000 --mixed 1000000 ! gets ! wall 0.621014 comb 0.620000 user 0.620000 sys 0.000000 (best of 16) ! wall 0.615146 comb 0.620000 user 0.620000 sys 0.000000 (best of 17) ! inserts ! <not available> ! wall 0.698115 comb 0.700000 user 0.700000 sys 0.000000 (best of 15) ! sets ! wall 0.560247 comb 0.560000 user 0.560000 sys 0.000000 (best of 18) ! wall 0.832495 comb 0.830000 user 0.830000 sys 0.000000 (best of 12) ! mixed ! wall 0.686172 comb 0.680000 user 0.680000 sys 0.000000 (best of 15) ! wall 0.841359 comb 0.840000 user 0.840000 sys 0.000000 (best of 12) We're still under 1us per insert, which seems like reasonable performance for a cache. If we comment out updating of self.totalcost during insert(), performance of insert() is identical to __setitem__ before. However, I don't want to make total cost evaluation lazy because it has significant performance implications for when we need to evaluate the total cost at mutation time (it requires a cache traversal, which could be expensive for large caches). Differential Revision: https://phab.mercurial-scm.org/D4502

util: add a popoldest() method to lrucachedict This allows consumers to prune the oldest item from the cache. This could be useful for e.g. a consumer that wishes for the size of items tracked by the cache to remain under a high water mark. Differential Revision: https://phab.mercurial-scm.org/D4501

util: ability to change capacity when copying lrucachedict This will allow us to easily replace an lrucachedict with one with a higher or lower capacity as consumers deem necessary. IMO it is easier to just create a new cache instance than to muck with the capacity of an existing cache. Mutating an existing cache's capacity feels more prone to bugs. Differential Revision: https://phab.mercurial-scm.org/D4500

util: make capacity a public attribute on lrucachedict So others can query it. Useful for operations that may want to verify the cache has capacity for N items before it performs an operation that may cause cache eviction. Differential Revision: https://phab.mercurial-scm.org/D4499

util: properly copy lrucachedict instances Previously, copy() only worked if the cache was full. We teach copy() to only copy defined nodes. Differential Revision: https://phab.mercurial-scm.org/D4498

tests: rewrite test-lrucachedict.py to use unittest This makes the code so much easier to test and debug. Along the way, I discovered a bug in copy(), which I kind of added test coverage for. Differential Revision: https://phab.mercurial-scm.org/D4497

wireprotov2peer: stream decoded responses Previously, wire protocol version 2 would buffer all response data. Only once all data was received did we CBOR decode it and resolve the future associated with the command. This was obviously not desirable. In future commits that introduce large response payloads, this caused significant memory bloat and slowed down client operations due to waiting on the server. This commit refactors the response handling code so that response data can be streamed. Command response objects now contain a buffered CBOR decoder. As new data arrives, it is fed into the decoder. Decoded objects are made available to the generator as they are decoded. Because there is a separate thread processing incoming frames and feeding data into the response object, there is the potential for race conditions when mutating response objects. So a lock has been added to guard access to critical state variables. Because the generator emitting decoded objects needs to wait on those objects to become available, we've added an Event for the generator to wait on so it doesn't busy loop. This does mean there is the potential for deadlocks. And I'm pretty sure they can occur in some scenarios. We already have a handful of TODOs around this. But I've added some more. Fixing this will likely require moving the background thread receiving frames into clienthandler. We likely would have done this anyway when implementing the client bits for the SSH transport. Test output changes because the initial CBOR map holding the overall response state is now always handled internally by the response object. Differential Revision: https://phab.mercurial-scm.org/D4474

wireprotoframing: buffer emitted data to reduce frame count An upcoming commit introduces a wire protocol command that can emit hundreds of thousands of small objects. Without a buffering layer, we would emit a single, small frame for every object. Performance profiling revealed this to be a source of significant overhead for both client and server. This commit introduces a very crude buffering layer so that we emit fewer, bigger frames in such a scenario. This code will likely get rewritten in the future to be part of the streams API, as we'll need a similar strategy for compressing data. I don't want to think about it too much at the moment though. server before: user 32.500+0.000 sys 1.160+0.000 after: user 20.230+0.010 sys 0.180+0.000 client before: user 133.400+0.000 sys 93.120+0.000 after: user 68.370+0.000 sys 32.950+0.000 This appears to indicate we have significant overhead in the frame processing code on both client and server. It might be worth profiling that at some point... Differential Revision: https://phab.mercurial-scm.org/D4473

wireprotov2: implement commands as a generator of objects Previously, wire protocol version 2 inherited version 1's model of having separate types to represent the results of different wire protocol commands. As I implemented more powerful commands in future commits, I found I was using a common pattern of returning a special type to hold a generator. This meant the command function required a closure to do most of the work. That made logic flow more difficult to follow. I also noticed that many commands were effectively a sequence of objects to be CBOR encoded. I think it makes sense to define version 2 commands as generators. This way, commands can simply emit the data structures they wish to send to the client. This eliminates the need for a closure in command functions and removes encoding from the bodies of commands. As part of this commit, the handling of response objects has been moved into the serverreactor class. This puts the reactor in the driver's seat with regards to CBOR encoding and error handling. Having error handling in the function that emits frames is particularly important because exceptions in that function can lead to things getting in a bad state: I'm fairly certain that uncaught exceptions in the frame generator were causing deadlocks. I also introduced a dedicated error type for explicit error reporting in command handlers. This will be used in subsequent commits. There's still a bit of work to be done here, especially around formalizing the error handling "protocol." I've added yet another TODO to track this so we don't forget. Test output changed because we're using generators and no longer know we are at the end of the data until we hit the end of the generator. This means we can't emit the end-of-stream flag until we've exhausted the generator. Hence the introduction of 0-sized end-of-stream frames. Differential Revision: https://phab.mercurial-scm.org/D4472