wireprotov2: support response caching
One of the things I've learned from managing VCS servers over the
years is that they are hard to scale. It is well known that some
companies have very beefy (read: very expensive) servers to power
their VCS needs. It is also known that specialized servers for
various VCS exist in order to facilitate scaling servers. (Mercurial
is in this boat.)
One of the aspects that make a VCS server hard to scale is the
high CPU load incurred by constant client clone/pull operations.
To alleviate the scaling pain associated with data retrieval
operations, I want to integrate caching into the Mercurial wire
protocol server as robustly as possible such that servers can
aggressively cache responses and defer as much server load as
possible.
This commit represents the initial implementation of a general
caching layer in wire protocol version 2.
We define a new interface and behavior for a wire protocol cacher
in repository.py. (This is probably where a reviewer should look
first to understand what is going on.)
The bulk of the added code is in wireprotov2server.py, where we
define how a command can opt in to being cached and integrate
caching into command dispatching.
From a very high-level:
* A command can declare itself as cacheable by providing a callable
that can be used to derive a cache key.
* At dispatch time, if a command is cacheable, we attempt to
construct a cacher and use it for serving the request and/or
caching the request.
* The dispatch layer handles the bulk of the business logic for
caching, making cachers mostly "dumb content stores."
* The mechanism for invalidating cached entries (one of the harder
parts about caching in general) is by varying the cache key when
state changes. As such, cachers don't need to be concerned with
cache invalidation.
Initially, we've hooked up support for caching "manifestdata" and
"filedata" commands. These are the simplest to cache, as they should
be immutable over time. Caching of commands related to changeset
data is a bit harder (because cache validation is impacted by
changes to bookmarks, phases, etc). This will be implemented later.
(Strictly speaking, censoring a file should invalidate caches. I've
added an inline TODO to track this edge case.)
To prove it works, this commit implements a test-only extension
providing in-memory caching backed by an lrucachedict. A new test
showing this extension behaving properly is added. FWIW, the
cacher is ~50 lines of code, demonstrating the relative ease with
which a cache can be added to a server.
While the test cacher is not suitable for production workloads, just
for kicks I performed a clone of just the changeset and manifest data
for the mozilla-unified repository. With a fully warmed cache (of just
the manifest data since changeset data is not cached), server-side
CPU usage dropped from ~73s to ~28s. That's pretty significant and
demonstrates the potential that response caching has on server
scalability!
Differential Revision: https://phab.mercurial-scm.org/D4773
# progress.py progress bars related code
#
# Copyright (C) 2010 Augie Fackler <durin42@gmail.com>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
from __future__ import absolute_import
import errno
import threading
import time
from .i18n import _
from . import encoding
def spacejoin(*args):
return ' '.join(s for s in args if s)
def shouldprint(ui):
return not (ui.quiet or ui.plain('progress')) and (
ui._isatty(ui.ferr) or ui.configbool('progress', 'assume-tty'))
def fmtremaining(seconds):
"""format a number of remaining seconds in human readable way
This will properly display seconds, minutes, hours, days if needed"""
if seconds < 60:
# i18n: format XX seconds as "XXs"
return _("%02ds") % (seconds)
minutes = seconds // 60
if minutes < 60:
seconds -= minutes * 60
# i18n: format X minutes and YY seconds as "XmYYs"
return _("%dm%02ds") % (minutes, seconds)
# we're going to ignore seconds in this case
minutes += 1
hours = minutes // 60
minutes -= hours * 60
if hours < 30:
# i18n: format X hours and YY minutes as "XhYYm"
return _("%dh%02dm") % (hours, minutes)
# we're going to ignore minutes in this case
hours += 1
days = hours // 24
hours -= days * 24
if days < 15:
# i18n: format X days and YY hours as "XdYYh"
return _("%dd%02dh") % (days, hours)
# we're going to ignore hours in this case
days += 1
weeks = days // 7
days -= weeks * 7
if weeks < 55:
# i18n: format X weeks and YY days as "XwYYd"
return _("%dw%02dd") % (weeks, days)
# we're going to ignore days and treat a year as 52 weeks
weeks += 1
years = weeks // 52
weeks -= years * 52
# i18n: format X years and YY weeks as "XyYYw"
return _("%dy%02dw") % (years, weeks)
# file_write() and file_flush() of Python 2 do not restart on EINTR if
# the file is attached to a "slow" device (e.g. a terminal) and raise
# IOError. We cannot know how many bytes would be written by file_write(),
# but a progress text is known to be short enough to be written by a
# single write() syscall, so we can just retry file_write() with the whole
# text. (issue5532)
#
# This should be a short-term workaround. We'll need to fix every occurrence
# of write() to a terminal or pipe.
def _eintrretry(func, *args):
while True:
try:
return func(*args)
except IOError as err:
if err.errno == errno.EINTR:
continue
raise
class progbar(object):
def __init__(self, ui):
self.ui = ui
self._refreshlock = threading.Lock()
self.resetstate()
def resetstate(self):
self.topics = []
self.topicstates = {}
self.starttimes = {}
self.startvals = {}
self.printed = False
self.lastprint = time.time() + float(self.ui.config(
'progress', 'delay'))
self.curtopic = None
self.lasttopic = None
self.indetcount = 0
self.refresh = float(self.ui.config(
'progress', 'refresh'))
self.changedelay = max(3 * self.refresh,
float(self.ui.config(
'progress', 'changedelay')))
self.order = self.ui.configlist('progress', 'format')
self.estimateinterval = self.ui.configwith(
float, 'progress', 'estimateinterval')
def show(self, now, topic, pos, item, unit, total):
if not shouldprint(self.ui):
return
termwidth = self.width()
self.printed = True
head = ''
needprogress = False
tail = ''
for indicator in self.order:
add = ''
if indicator == 'topic':
add = topic
elif indicator == 'number':
if total:
add = b'%*d/%d' % (len(str(total)), pos, total)
else:
add = b'%d' % pos
elif indicator.startswith('item') and item:
slice = 'end'
if '-' in indicator:
wid = int(indicator.split('-')[1])
elif '+' in indicator:
slice = 'beginning'
wid = int(indicator.split('+')[1])
else:
wid = 20
if slice == 'end':
add = encoding.trim(item, wid, leftside=True)
else:
add = encoding.trim(item, wid)
add += (wid - encoding.colwidth(add)) * ' '
elif indicator == 'bar':
add = ''
needprogress = True
elif indicator == 'unit' and unit:
add = unit
elif indicator == 'estimate':
add = self.estimate(topic, pos, total, now)
elif indicator == 'speed':
add = self.speed(topic, pos, unit, now)
if not needprogress:
head = spacejoin(head, add)
else:
tail = spacejoin(tail, add)
if needprogress:
used = 0
if head:
used += encoding.colwidth(head) + 1
if tail:
used += encoding.colwidth(tail) + 1
progwidth = termwidth - used - 3
if total and pos <= total:
amt = pos * progwidth // total
bar = '=' * (amt - 1)
if amt > 0:
bar += '>'
bar += ' ' * (progwidth - amt)
else:
progwidth -= 3
self.indetcount += 1
# mod the count by twice the width so we can make the
# cursor bounce between the right and left sides
amt = self.indetcount % (2 * progwidth)
amt -= progwidth
bar = (' ' * int(progwidth - abs(amt)) + '<=>' +
' ' * int(abs(amt)))
prog = ''.join(('[', bar, ']'))
out = spacejoin(head, prog, tail)
else:
out = spacejoin(head, tail)
self._writeerr('\r' + encoding.trim(out, termwidth))
self.lasttopic = topic
self._flusherr()
def clear(self):
if not self.printed or not self.lastprint or not shouldprint(self.ui):
return
self._writeerr('\r%s\r' % (' ' * self.width()))
if self.printed:
# force immediate re-paint of progress bar
self.lastprint = 0
def complete(self):
if not shouldprint(self.ui):
return
if self.ui.configbool('progress', 'clear-complete'):
self.clear()
else:
self._writeerr('\n')
self._flusherr()
def _flusherr(self):
_eintrretry(self.ui.ferr.flush)
def _writeerr(self, msg):
_eintrretry(self.ui.ferr.write, msg)
def width(self):
tw = self.ui.termwidth()
return min(int(self.ui.config('progress', 'width', default=tw)), tw)
def estimate(self, topic, pos, total, now):
if total is None:
return ''
initialpos = self.startvals[topic]
target = total - initialpos
delta = pos - initialpos
if delta > 0:
elapsed = now - self.starttimes[topic]
seconds = (elapsed * (target - delta)) // delta + 1
return fmtremaining(seconds)
return ''
def speed(self, topic, pos, unit, now):
initialpos = self.startvals[topic]
delta = pos - initialpos
elapsed = now - self.starttimes[topic]
if elapsed > 0:
return _('%d %s/sec') % (delta / elapsed, unit)
return ''
def _oktoprint(self, now):
'''Check if conditions are met to print - e.g. changedelay elapsed'''
if (self.lasttopic is None # first time we printed
# not a topic change
or self.curtopic == self.lasttopic
# it's been long enough we should print anyway
or now - self.lastprint >= self.changedelay):
return True
else:
return False
def _calibrateestimate(self, topic, now, pos):
'''Adjust starttimes and startvals for topic so ETA works better
If progress is non-linear (ex. get much slower in the last minute),
it's more friendly to only use a recent time span for ETA and speed
calculation.
[======================================> ]
^^^^^^^
estimateinterval, only use this for estimation
'''
interval = self.estimateinterval
if interval <= 0:
return
elapsed = now - self.starttimes[topic]
if elapsed > interval:
delta = pos - self.startvals[topic]
newdelta = delta * interval / elapsed
# If a stall happens temporarily, ETA could change dramatically
# frequently. This is to avoid such dramatical change and make ETA
# smoother.
if newdelta < 0.1:
return
self.startvals[topic] = pos - newdelta
self.starttimes[topic] = now - interval
def progress(self, topic, pos, item='', unit='', total=None):
if pos is None:
self.closetopic(topic)
return
now = time.time()
with self._refreshlock:
if topic not in self.topics:
self.starttimes[topic] = now
self.startvals[topic] = pos
self.topics.append(topic)
self.topicstates[topic] = pos, item, unit, total
self.curtopic = topic
self._calibrateestimate(topic, now, pos)
if now - self.lastprint >= self.refresh and self.topics:
if self._oktoprint(now):
self.lastprint = now
self.show(now, topic, *self.topicstates[topic])
def closetopic(self, topic):
with self._refreshlock:
self.starttimes.pop(topic, None)
self.startvals.pop(topic, None)
self.topicstates.pop(topic, None)
# reset the progress bar if this is the outermost topic
if self.topics and self.topics[0] == topic and self.printed:
self.complete()
self.resetstate()
# truncate the list of topics assuming all topics within
# this one are also closed
if topic in self.topics:
self.topics = self.topics[:self.topics.index(topic)]
# reset the last topic to the one we just unwound to,
# so that higher-level topics will be stickier than
# lower-level topics
if self.topics:
self.lasttopic = self.topics[-1]
else:
self.lasttopic = None