view contrib/synthrepo.py @ 40021:c537144fdbef

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

# synthrepo.py - repo synthesis
#
# Copyright 2012 Facebook
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.

'''synthesize structurally interesting change history

This extension is useful for creating a repository with properties
that are statistically similar to an existing repository. During
analysis, a simple probability table is constructed from the history
of an existing repository.  During synthesis, these properties are
reconstructed.

Properties that are analyzed and synthesized include the following:

- Lines added or removed when an existing file is modified
- Number and sizes of files added
- Number of files removed
- Line lengths
- Topological distance to parent changeset(s)
- Probability of a commit being a merge
- Probability of a newly added file being added to a new directory
- Interarrival time, and time zone, of commits
- Number of files in each directory

A few obvious properties that are not currently handled realistically:

- Merges are treated as regular commits with two parents, which is not
  realistic
- Modifications are not treated as operations on hunks of lines, but
  as insertions and deletions of randomly chosen single lines
- Committer ID (always random)
- Executability of files
- Symlinks and binary files are ignored
'''

from __future__ import absolute_import
import bisect
import collections
import itertools
import json
import os
import random
import sys
import time

from mercurial.i18n import _
from mercurial.node import (
    nullid,
    nullrev,
    short,
)
from mercurial import (
    context,
    diffutil,
    error,
    hg,
    patch,
    registrar,
    scmutil,
)
from mercurial.utils import (
    dateutil,
)

# Note for extension authors: ONLY specify testedwith = 'ships-with-hg-core' for
# extensions which SHIP WITH MERCURIAL. Non-mainline extensions should
# be specifying the version(s) of Mercurial they are tested with, or
# leave the attribute unspecified.
testedwith = 'ships-with-hg-core'

cmdtable = {}
command = registrar.command(cmdtable)

newfile = {'new fi', 'rename', 'copy f', 'copy t'}

def zerodict():
    return collections.defaultdict(lambda: 0)

def roundto(x, k):
    if x > k * 2:
        return int(round(x / float(k)) * k)
    return int(round(x))

def parsegitdiff(lines):
    filename, mar, lineadd, lineremove = None, None, zerodict(), 0
    binary = False
    for line in lines:
        start = line[:6]
        if start == 'diff -':
            if filename:
                yield filename, mar, lineadd, lineremove, binary
            mar, lineadd, lineremove, binary = 'm', zerodict(), 0, False
            filename = patch.gitre.match(line).group(1)
        elif start in newfile:
            mar = 'a'
        elif start == 'GIT bi':
            binary = True
        elif start == 'delete':
            mar = 'r'
        elif start:
            s = start[0]
            if s == '-' and not line.startswith('--- '):
                lineremove += 1
            elif s == '+' and not line.startswith('+++ '):
                lineadd[roundto(len(line) - 1, 5)] += 1
    if filename:
        yield filename, mar, lineadd, lineremove, binary

@command('analyze',
         [('o', 'output', '', _('write output to given file'), _('FILE')),
          ('r', 'rev', [], _('analyze specified revisions'), _('REV'))],
         _('hg analyze'), optionalrepo=True)
def analyze(ui, repo, *revs, **opts):
    '''create a simple model of a repository to use for later synthesis

    This command examines every changeset in the given range (or all
    of history if none are specified) and creates a simple statistical
    model of the history of the repository. It also measures the directory
    structure of the repository as checked out.

    The model is written out to a JSON file, and can be used by
    :hg:`synthesize` to create or augment a repository with synthetic
    commits that have a structure that is statistically similar to the
    analyzed repository.
    '''
    root = repo.root
    if not root.endswith(os.path.sep):
        root += os.path.sep

    revs = list(revs)
    revs.extend(opts['rev'])
    if not revs:
        revs = [':']

    output = opts['output']
    if not output:
        output = os.path.basename(root) + '.json'

    if output == '-':
        fp = sys.stdout
    else:
        fp = open(output, 'w')

    # Always obtain file counts of each directory in the given root directory.
    def onerror(e):
        ui.warn(_('error walking directory structure: %s\n') % e)

    dirs = {}
    rootprefixlen = len(root)
    for dirpath, dirnames, filenames in os.walk(root, onerror=onerror):
        dirpathfromroot = dirpath[rootprefixlen:]
        dirs[dirpathfromroot] = len(filenames)
        if '.hg' in dirnames:
            dirnames.remove('.hg')

    lineschanged = zerodict()
    children = zerodict()
    p1distance = zerodict()
    p2distance = zerodict()
    linesinfilesadded = zerodict()
    fileschanged = zerodict()
    filesadded = zerodict()
    filesremoved = zerodict()
    linelengths = zerodict()
    interarrival = zerodict()
    parents = zerodict()
    dirsadded = zerodict()
    tzoffset = zerodict()

    # If a mercurial repo is available, also model the commit history.
    if repo:
        revs = scmutil.revrange(repo, revs)
        revs.sort()

        progress = ui.makeprogress(_('analyzing'), unit=_('changesets'),
                                   total=len(revs))
        for i, rev in enumerate(revs):
            progress.update(i)
            ctx = repo[rev]
            pl = ctx.parents()
            pctx = pl[0]
            prev = pctx.rev()
            children[prev] += 1
            p1distance[rev - prev] += 1
            parents[len(pl)] += 1
            tzoffset[ctx.date()[1]] += 1
            if len(pl) > 1:
                p2distance[rev - pl[1].rev()] += 1
            if prev == rev - 1:
                lastctx = pctx
            else:
                lastctx = repo[rev - 1]
            if lastctx.rev() != nullrev:
                timedelta = ctx.date()[0] - lastctx.date()[0]
                interarrival[roundto(timedelta, 300)] += 1
            diffopts = diffutil.diffallopts(ui, {'git': True})
            diff = sum((d.splitlines()
                       for d in ctx.diff(pctx, opts=diffopts)), [])
            fileadds, diradds, fileremoves, filechanges = 0, 0, 0, 0
            for filename, mar, lineadd, lineremove, isbin in parsegitdiff(diff):
                if isbin:
                    continue
                added = sum(lineadd.itervalues(), 0)
                if mar == 'm':
                    if added and lineremove:
                        lineschanged[roundto(added, 5),
                                     roundto(lineremove, 5)] += 1
                        filechanges += 1
                elif mar == 'a':
                    fileadds += 1
                    if '/' in filename:
                        filedir = filename.rsplit('/', 1)[0]
                        if filedir not in pctx.dirs():
                            diradds += 1
                    linesinfilesadded[roundto(added, 5)] += 1
                elif mar == 'r':
                    fileremoves += 1
                for length, count in lineadd.iteritems():
                    linelengths[length] += count
            fileschanged[filechanges] += 1
            filesadded[fileadds] += 1
            dirsadded[diradds] += 1
            filesremoved[fileremoves] += 1
        progress.complete()

    invchildren = zerodict()

    for rev, count in children.iteritems():
        invchildren[count] += 1

    if output != '-':
        ui.status(_('writing output to %s\n') % output)

    def pronk(d):
        return sorted(d.iteritems(), key=lambda x: x[1], reverse=True)

    json.dump({'revs': len(revs),
               'initdirs': pronk(dirs),
               'lineschanged': pronk(lineschanged),
               'children': pronk(invchildren),
               'fileschanged': pronk(fileschanged),
               'filesadded': pronk(filesadded),
               'linesinfilesadded': pronk(linesinfilesadded),
               'dirsadded': pronk(dirsadded),
               'filesremoved': pronk(filesremoved),
               'linelengths': pronk(linelengths),
               'parents': pronk(parents),
               'p1distance': pronk(p1distance),
               'p2distance': pronk(p2distance),
               'interarrival': pronk(interarrival),
               'tzoffset': pronk(tzoffset),
               },
              fp)
    fp.close()

@command('synthesize',
         [('c', 'count', 0, _('create given number of commits'), _('COUNT')),
          ('', 'dict', '', _('path to a dictionary of words'), _('FILE')),
          ('', 'initfiles', 0, _('initial file count to create'), _('COUNT'))],
         _('hg synthesize [OPTION].. DESCFILE'))
def synthesize(ui, repo, descpath, **opts):
    '''synthesize commits based on a model of an existing repository

    The model must have been generated by :hg:`analyze`. Commits will
    be generated randomly according to the probabilities described in
    the model. If --initfiles is set, the repository will be seeded with
    the given number files following the modeled repository's directory
    structure.

    When synthesizing new content, commit descriptions, and user
    names, words will be chosen randomly from a dictionary that is
    presumed to contain one word per line. Use --dict to specify the
    path to an alternate dictionary to use.
    '''
    try:
        fp = hg.openpath(ui, descpath)
    except Exception as err:
        raise error.Abort('%s: %s' % (descpath, err[0].strerror))
    desc = json.load(fp)
    fp.close()

    def cdf(l):
        if not l:
            return [], []
        vals, probs = zip(*sorted(l, key=lambda x: x[1], reverse=True))
        t = float(sum(probs, 0))
        s, cdfs = 0, []
        for v in probs:
            s += v
            cdfs.append(s / t)
        return vals, cdfs

    lineschanged = cdf(desc['lineschanged'])
    fileschanged = cdf(desc['fileschanged'])
    filesadded = cdf(desc['filesadded'])
    dirsadded = cdf(desc['dirsadded'])
    filesremoved = cdf(desc['filesremoved'])
    linelengths = cdf(desc['linelengths'])
    parents = cdf(desc['parents'])
    p1distance = cdf(desc['p1distance'])
    p2distance = cdf(desc['p2distance'])
    interarrival = cdf(desc['interarrival'])
    linesinfilesadded = cdf(desc['linesinfilesadded'])
    tzoffset = cdf(desc['tzoffset'])

    dictfile = opts.get('dict') or '/usr/share/dict/words'
    try:
        fp = open(dictfile, 'rU')
    except IOError as err:
        raise error.Abort('%s: %s' % (dictfile, err.strerror))
    words = fp.read().splitlines()
    fp.close()

    initdirs = {}
    if desc['initdirs']:
        for k, v in desc['initdirs']:
            initdirs[k.encode('utf-8').replace('.hg', '_hg')] = v
        initdirs = renamedirs(initdirs, words)
    initdirscdf = cdf(initdirs)

    def pick(cdf):
        return cdf[0][bisect.bisect_left(cdf[1], random.random())]

    def pickpath():
        return os.path.join(pick(initdirscdf), random.choice(words))

    def makeline(minimum=0):
        total = max(minimum, pick(linelengths))
        c, l = 0, []
        while c < total:
            w = random.choice(words)
            c += len(w) + 1
            l.append(w)
        return ' '.join(l)

    wlock = repo.wlock()
    lock = repo.lock()

    nevertouch = {'.hgsub', '.hgignore', '.hgtags'}

    _synthesizing = _('synthesizing')
    _files = _('initial files')
    _changesets = _('changesets')

    # Synthesize a single initial revision adding files to the repo according
    # to the modeled directory structure.
    initcount = int(opts['initfiles'])
    if initcount and initdirs:
        pctx = repo[None].parents()[0]
        dirs = set(pctx.dirs())
        files = {}

        def validpath(path):
            # Don't pick filenames which are already directory names.
            if path in dirs:
                return False
            # Don't pick directories which were used as file names.
            while path:
                if path in files:
                    return False
                path = os.path.dirname(path)
            return True

        progress = ui.makeprogress(_synthesizing, unit=_files, total=initcount)
        for i in xrange(0, initcount):
            progress.update(i)

            path = pickpath()
            while not validpath(path):
                path = pickpath()
            data = '%s contents\n' % path
            files[path] = data
            dir = os.path.dirname(path)
            while dir and dir not in dirs:
                dirs.add(dir)
                dir = os.path.dirname(dir)

        def filectxfn(repo, memctx, path):
            return context.memfilectx(repo, memctx, path, files[path])

        progress.complete()
        message = 'synthesized wide repo with %d files' % (len(files),)
        mc = context.memctx(repo, [pctx.node(), nullid], message,
                            files, filectxfn, ui.username(),
                            '%d %d' % dateutil.makedate())
        initnode = mc.commit()
        if ui.debugflag:
            hexfn = hex
        else:
            hexfn = short
        ui.status(_('added commit %s with %d files\n')
                  % (hexfn(initnode), len(files)))

    # Synthesize incremental revisions to the repository, adding repo depth.
    count = int(opts['count'])
    heads = set(map(repo.changelog.rev, repo.heads()))
    progress = ui.makeprogress(_synthesizing, unit=_changesets, total=count)
    for i in xrange(count):
        progress.update(i)

        node = repo.changelog.node
        revs = len(repo)

        def pickhead(heads, distance):
            if heads:
                lheads = sorted(heads)
                rev = revs - min(pick(distance), revs)
                if rev < lheads[-1]:
                    rev = lheads[bisect.bisect_left(lheads, rev)]
                else:
                    rev = lheads[-1]
                return rev, node(rev)
            return nullrev, nullid

        r1 = revs - min(pick(p1distance), revs)
        p1 = node(r1)

        # the number of heads will grow without bound if we use a pure
        # model, so artificially constrain their proliferation
        toomanyheads = len(heads) > random.randint(1, 20)
        if p2distance[0] and (pick(parents) == 2 or toomanyheads):
            r2, p2 = pickhead(heads.difference([r1]), p2distance)
        else:
            r2, p2 = nullrev, nullid

        pl = [p1, p2]
        pctx = repo[r1]
        mf = pctx.manifest()
        mfk = mf.keys()
        changes = {}
        if mfk:
            for __ in xrange(pick(fileschanged)):
                for __ in xrange(10):
                    fctx = pctx.filectx(random.choice(mfk))
                    path = fctx.path()
                    if not (path in nevertouch or fctx.isbinary() or
                            'l' in fctx.flags()):
                        break
                lines = fctx.data().splitlines()
                add, remove = pick(lineschanged)
                for __ in xrange(remove):
                    if not lines:
                        break
                    del lines[random.randrange(0, len(lines))]
                for __ in xrange(add):
                    lines.insert(random.randint(0, len(lines)), makeline())
                path = fctx.path()
                changes[path] = '\n'.join(lines) + '\n'
            for __ in xrange(pick(filesremoved)):
                path = random.choice(mfk)
                for __ in xrange(10):
                    path = random.choice(mfk)
                    if path not in changes:
                        break
        if filesadded:
            dirs = list(pctx.dirs())
            dirs.insert(0, '')
        for __ in xrange(pick(filesadded)):
            pathstr = ''
            while pathstr in dirs:
                path = [random.choice(dirs)]
                if pick(dirsadded):
                    path.append(random.choice(words))
                path.append(random.choice(words))
                pathstr = '/'.join(filter(None, path))
            data = '\n'.join(makeline()
                             for __ in xrange(pick(linesinfilesadded))) + '\n'
            changes[pathstr] = data
        def filectxfn(repo, memctx, path):
            if path not in changes:
                return None
            return context.memfilectx(repo, memctx, path, changes[path])
        if not changes:
            continue
        if revs:
            date = repo['tip'].date()[0] + pick(interarrival)
        else:
            date = time.time() - (86400 * count)
        # dates in mercurial must be positive, fit in 32-bit signed integers.
        date = min(0x7fffffff, max(0, date))
        user = random.choice(words) + '@' + random.choice(words)
        mc = context.memctx(repo, pl, makeline(minimum=2),
                            sorted(changes),
                            filectxfn, user, '%d %d' % (date, pick(tzoffset)))
        newnode = mc.commit()
        heads.add(repo.changelog.rev(newnode))
        heads.discard(r1)
        heads.discard(r2)
    progress.complete()

    lock.release()
    wlock.release()

def renamedirs(dirs, words):
    '''Randomly rename the directory names in the per-dir file count dict.'''
    wordgen = itertools.cycle(words)
    replacements = {'': ''}
    def rename(dirpath):
        '''Recursively rename the directory and all path prefixes.

        The mapping from path to renamed path is stored for all path prefixes
        as in dynamic programming, ensuring linear runtime and consistent
        renaming regardless of iteration order through the model.
        '''
        if dirpath in replacements:
            return replacements[dirpath]
        head, _ = os.path.split(dirpath)
        if head:
            head = rename(head)
        else:
            head = ''
        renamed = os.path.join(head, next(wordgen))
        replacements[dirpath] = renamed
        return renamed
    result = []
    for dirpath, count in dirs.iteritems():
        result.append([rename(dirpath.lstrip(os.sep)), count])
    return result