Re: [RFD] Gitweb caching, part 1 (long)

[Please Cc: me directly, as I am not subscribed to git mailing list,
 and GMane NNTP (news) interface I use doesn't show currently any new
 posts; I wouldn't want to miss any response.  Thanks in advance.]

This post shows my ideas about how to implement caching in gitweb, my thoughts on what are the problems, and what solutions (what code) can we use.

bottlenecks on git mailing list, the main culprit is projects list (which perhaps should be redesigned), and summary pages for some of larger / more popular projects. Gitweb performance is I/O bound, not CPU bound, so I guess not all existing caching solutions and ideas would work with gitweb.

There are some troubles with adding generic (as opposed to site-specific) caching to gitweb. First, gitweb should work both with mod_perl and as CGI (perhaps in the future also FastCGI) script. Second, the solution should not depend on additional packages, at least not those that can be found packaged in extras or well trusted contrib repositories; not all admins allow installing packages from CPAN. Third, the solution could be helped but should not depend on adding helper hooks to users repositories; while hosting sites like repo.or.cz controls repositories, sites like kernel.org or freedesktop.org, which give shell access, do not.

Let talk first about what to cache.

1. Support for caching in HTTP (HTTP accelerators, caching engines)

My first idea of adding caching support to gitweb was for it to
generate proper "freshness" caching headers (Last-Modified: and ETag:)
and respond to cache validation requests (If-Modified-Since:,
If-None-Match: etc.), and for reverse proxy, aka. caching engine,
aka. web accelerator/HTTP accelerator (e.g. Varnish or Squid) take
care of caching.  It is better to use existing solution, isn't it?

Unfortunately gitweb to generate date for If-Modified-Since:, or a tag for If-None-Match: must do hard work; perhaps not as much as generating the whole page in the term of CPU, but almost the same in the terms of I/O hit. So it is not so simple...

Nevertheless even if using reverse proxy for gitweb caching is not so simple, gitweb should play well with support for caching in HTTP protocol, so the pages can be cached between gitweb and user, either in one of intermediate caches, proxy server with caching support, or browser cache. Currently gitweb uses 'Expires:' header with expiry of 24h / 1d (IIRC cutoff time for caches; also IIRC forever is half a year) for pages which we know would not change (using full SHA-1 identifier and all required information filled). We should probably generate Last-Modified: and/or ETag: if it is possible.

However if gitweb has some kind of internal caching turned on, it can respond properly to validation requests with low cost. This way some of requests would be handled by intermediate caches, so gitweb wouldn't have even to access the cache to return an answer. But IMHO that is a secondary concern: it could help, but isn't possible to do well without in-gitweb caching (as far as I can see).

BTW. besides optionally marking result as being retrieved from cache
("stale" or "cached"), gitweb I think should also send appropriate
Warning: header, see sections 13.1.2 and 14.46 of RFC2616, e.g.
  Warning: 110 git.kernel.org "Response is stale"

References:
* "Caching Tutorial for Web Authors"
  http://www.mnot.net/cache_docs/
* HTTP 1.1 Specification (RFC 2616)
  http://www.ietf.org/rfc/rfc2616.txt

2. Caching Perl structures

On of solutions (used for example by Petr 'Pasky' Baudis in his last post about caching projects list info in gitweb) is be to cache (save) Perl structures containing information needed to generate response (web page). Another solution, discussed below, would be to cache generated output, i.e. web page, optionally with (some) HTTP headers.

The advantage of storing Perl structures (raw data) in the cache is that the same data can be reused for different pages (e.g. paging projects list if/when it gets implemented), same page with varying part (e.g. content type being text/html or application/xhtml+xml depending on what web browser prefers, or transparently compressed output via Transfer-Encoding: depending on web browser capabilities), and for replying to cache validation requests. Additionally we can generate web pages with correct relative (e.g. "5 minutes ago") time info. Not that all but first and last are not possible with caching [final] output, but it would be, I think, much harder...

The disadvantage is that we have to decide is what format to use for serializing data, i.e. to represent compound complex data as stream of bytes in cache... unless of course gitweb would rely on one of already existing caching solutions, which usually take care of this problem for us: see next section (in next installment).

Formats I was considering were:
 - Data::Dumper
 - Storable (binary)
 - YAML Tiny
 - gitconfig tiny

2.1. Data::Dumper

One of advantages of Data::Dumper format is that it comes with Perl installation, so there is no problem with installing it (well, at least it comes with perl-5.8.6-24 RPM on Linux). Another advantage is that it is textual format, thus easy to debug in the case of problems.

Main problem with Data::Dumper is that it use eval() to thaw (restore) data from serialized form in cache, which is serious security risk in less secure environments.

2.2. Storable

Also comes packaged with Perl distribution. Offers writing to and restoring from file or to/from opened file handle. It is fast, from my unscientific tests around 3-4 times as fast as using eval() to read Data::Dumper data.

One of disadvantages is the fact that Storable format is binary, so
you would have to write separate Perl script to convert it to
human-readable form (e.g. Data::Dumper form).  Format includes format
and version header, and modern 'file' installations should detect it
correctly, e.g.:
  filename: perl Storable(v0.7) data (major 2) (minor 6)

Another nuisance is the fact that while Storable(3pm) manpage states that:

     The [retrieve()] routine returns "undef" for I/O problems or
     other internal error, a true value otherwise. Serious errors are
     propagated as a "die" exception.

But 'serious errors' include the fact that file is not in correct format, so for safety (because server-side script should return page with error info instead of dying silently) one would have to use "eval { ... }" to catch errors.

Used by Cache::Cache and, I think, also by other caching solutions (packages).

2.3. YAML Tiny (subset of YAML)

YAML was created as human-readable serialization format, easy to parse by machine. Unfortunately none of YAML parsing modules (YAML, YAML::Syck, YAML::Tiny) are packaged with Perl; on the other hand they can be found in Dries RPM repository, so I guess they fill criteria of being in extras or trusted contrib package repository.

Additionally at least YAML::Tiny (which implements subset of YAML in pure Perl code) is slower, around 4-5 times, than even using Data::Dumper, and more that 10 times slower than using Storable. This _might_ have been caused by the choice of module to implement YAML parsing.

We could write parser (and generator) for even smaller subset of YAML to use only those features that are truly needed by gitweb... but then we can go with the next format, which is also text format, and also doesn't have insecurities of using eval() to thaw data (read from cache).

YAML was designed from the ground up to be an excellent syntax for configuration files. Not necessarily so for cache.

References:
* http://en.wikipedia.org/wiki/YAML
* YAML::Tiny(3pm)
  http://search.cpan.org/dist/YAML-Tiny/lib/YAML/Tiny.pm
* YAML Ain't Markup Language (YAML^TM) Version 1.1, Working Draft
  http://yaml.org/spec/current.html

2.4. gitconfig tiny (subset of ini-like gitconfig format).

What, I think, we would want to cache is usually list of records, or in Perl terminology array of hashes; usually ordering of array doesn't matter.

Because of that I think it would be possible to represent data to be saved (cached) in the ini-like extended git config format. Then gitweb could either (re)use config parser in Perl used by git-cvsserver (which accepts subset of valid config format), or "git config --file=<path> -z -l" to slurp data in more parseable format... but if we do that, we could choose this format or variation of it as our serialization format.

The cache file could look like this:

   [gitweb "<primary key value>"]
   	key1 = value1
   	key2 = "value with spaces"

where for list of projects info primary key might be path (relative to projectroot) to the repository.

3. Caching output: formatted pages

Alternate solution to caching Perl structures is caching final output, with or without (some/all) HTTP headers. It has the advantage that it is simple to implement, and that the same code can be used to cache all the pages. (But we could get similar result by creating something similar to Tie::Memoize, tying hash or array so it automatically get data either from git command, or from cache... or we can implement universal API, like Cache::Cache API.)

This is from what I understand what kernel.org (warthog9) gitweb uses; I don't know what cgit (web interface in C) which also has some caching support uses: does it cache data or output?

How one can simply extend CGI script with support for caching is shown by the CGI::Cache (non-standard CPAN Perl module). On the other hand gitweb can afford more extensive surgery.

References:
* CGI::Cache(3pm)
  http://search.cpan.org/~dcoppit/CGI-Cache-1.4200/lib/CGI/Cache.pm

To be continued...

%% .................................................................. %% In next parts:

Cache lifetime and invalidation
1. static cache, external refreshing e.g. by hooks
2. stat and/or inotify, to check if repository changed
3. cache lifetime (trying to avoid "thundering horde" problem)

CPAN packages we could use, or take inspiration from
1. Cache::Cache (standard)
2. CHI, Unified caching interface
3. Cache
4. other (e.g. Cache::Adaptive, using Cache::Cache)

On Wed, Mar 19, 2008 at 01:54:53AM +0100, Jakub Narebski wrote:

git-cvsserver uses "git config -l", too.

Gruesse,

In previous part:

What to cache.
1. Support for caching in HTTP (external caching)
2. Caching Perl structures (and serialization)
3. Caching gitweb output: formatted pages

TO WHOM IT MAY CONCERN: John 'Warthog9' Hawley (J.H.) who created caching for gitweb at kernel.org; Petr 'Pasky' Baudis who maintains repo.or.cz fork of gitweb and lately added caching of projects list info; Lars Hjemli who is the author of cgit, git web interface in C which includes some caching. (BTW. I'd like to hear your thoughs on git web interface caching, and about solutions you have implemented).

This is continuation of my thoughts about how to implement caching in gitweb, what problems we could encounter, and what existing solutions (what code/what packages) can we (re)use.

One of the more important issues to think about when implementing caching is to decide when to regenerate cache, i.e. issues of cache (in)validation and lifetime.

1. Static cache, external refreshing (invalidation).

The easiest situation is when cache can be invalidated (removed) externally; caching support in gitweb would then need only to either use cached information or cached output if it exists, and generate information and/or output and cache appropriate things if cache doesn't exist.

In closed-up git hosting system like repo.or.cz new contents can appear in repository only via push (if repo is manually updated) or via automated fetch (if repo is mirrored automatically). This means that it is known when infomration about given repository gets stale (out of sync). It would be then enought to make 'update' or 'post-receive' hook to delete cache, or invalidate parts of cached info about given repository. Creation and deletion of repositories should also be handled by scripts; they affect caching too.

This of course assumes that we can control repository hooks (perhaps git should learn hook multiplexing first, as proposed some time ago on a mailing list). This is not the case when developers are given shell access, and gitweb is offered as a part of service rather than as a part of git hosting; repositories are not under web administrator control. This is the case (according to J.H. on git mailing list) for kernel.org.

So we have to examine also more generic solutions.

2. Checking filesystem (stat and/or inotify).

If new objects come to repository via commit or via fetch it is enough (I guess) to watch for modifications of GIT_DIR of a project (I think due to doing atomic writes via "create temporary file, then rename it to final filename" of files in GIT_DIR: COMMIT_MSG and FETCH_HEAD). So it should be enough to check and compare stat info for GIT_DIR of a project, or of possible implement some inotify (or equivalent on other operating systems than Linux) checking, to see if cache can contain stale info. In practice what we can truly check is that nothing changed with repo.

Unfortunately the above is not the case if objects come to repository via push. Note that both push resulting in crating a pack (this I think could use the same mechanism, only checking GIT_DIR/objects/pack directory), and push resulting in creation of loose objects has to be supported; additionally the refs pushed can have deeply hierarchical names.

I would be grateful if somebody could think a way to check if anything could have changed for such situation... but as it is now we have to go to more complicated ways of cache invalidation.

3. Cache lifetime.

Finally, for cases such as gitweb where validating cache (checking if the cached information isn't stale, out of sync with reality) is almost as costly as calculating the whole information without using cache at all, there is one possible solution to cache validation: simply keep cache for some time. For longer cache lifetimes gitweb perhaps should put some notification that information is from cache, perhaps with the time in human readable form how much time ago was this information generated (human readable means no "1325 seconds ago" info ;-). And if we want to be thorough, put it also in the HTTP header "Warning:" (at least for HTTP/1.1, see sections 13.1.2 and 14.46 of RFC 2616), e.g.:

  Warning: 110 git.kernel.org "Response is stale"

The question is what timeout, or how to choose lifetime of a cache. J.H. kernel.org's gitweb tries to adjust cache lifetime to server load, making cache lifetime longer if server load is higher, but ensuring that cache lifetime stays within specified bounds.

I have found among CPAN modules Cache::Adaptive where you can also specify bounds for expire time and subroutine to adjust cache lifetime, e.g. according to load average, process time for building the cache entry, etc. (it can use specified backend, for example Cache::FileCache from Cache::Cache distribution). Its subclass Cache::Adaptive::ByLoad which tries to adjust cache lifetime for bottlenecks under heavy load. Neither of modules I think is distributed as ready package in extras on trusted contrib packages repositories. Nevertheless we can "borrow" the algorithms used by those modules.

We should also try to avoid 'thundering herd' problem, namely that cache expires, gitweb gets N requests before cache gets re-created, and [poorly designed] cache architecture makes all N do the work regenerating cache. There are several ideas of how to deal with this problem:

 * If (part of) cache has expired, set its expiration time to the
   current time plus specified duration (slop) needed to regenerate
   cache.  It was used by original Pasky solution (and is used by
   further solutions for caching projects list sent here); in can be
   used by CHI (caching infrastructure) with busy_lock option... well,
   kind of.

 * Use some kind of locking so only one process does the work and
   updates the cache.  From what I've briefly checked that is what
   kernel.org gitweb does (using flock()).

   The patch implementing projects list info caching does protect
   using O_EXCL on temporary/lock file against more than one process
   writing the cache, but doesn't protect against more than one
   process doing the work, unformtunately.

 * Allows items to expire a little earlier than the stated expiration
   time to help prevent cache miss stampedes.  This is what CHI module
   does with expires_variance option.

   The probability of expiration increases as a function of how far
   along we are in the potential expiration window, with the
   probability being near 0 at the beginning of the window and
   approaching 1 at the end.

If cache size becomes issue there will be additional complications like which entries (which cached values) to remove first when we go over the cache size limit; but lets us leave it for later, if it would be needed at all.

%% In next part:

CPAN packages we could use, or take inspiration from
1. Cache::Cache (standard)
2. CHI - Unified cache interface
3. Cache - the Cache interface 
4. other interesting packages
  * Cache::Adaptive for adaptive cache lifetime solutions
  * Cache::Memcached and/or Cache::Swifty
    for caching using cache daemon 
  * Cache::FastMmap (also example of callbacks),
    and caching benchmark mentioned there

[Hopefully this resend wouldn't be stopped by vger antispam filter]

In previous parts:

What to cache.
1. Support for caching in HTTP (external caching).
2. Caching Perl structures (and serialization).
3. Caching gitweb output: formatted pages.

Cache (in)validation and lifetime
1. Static cache, external refreshing (invalidation).
2. Checking filesystem (stat and/or inotify).
3. Cache lifetime (timing-out cached info).
4. LRU (Least Recently Used) and others <- to be written

In this part I will write about existing caching solutions, or to be more exact about CPAN packages implementing caching or cache interface in Perl.

Note that not on all sites one can install packages from CPAN; only packages from main packages repository, from extras repository, or sometimes from trusted contrib repository are possible (see J.H. post about this problem). From mentioned packages only Cache::Cache and Cache::Mmap are available in contrib repositories for Aurox 11.1 distribution (based on Fedora Core 4): Cache::Cache distribution as perl-Cache-Cache in Dries RPM Repository, and Cache::Mmap as perl-Cache-Mmap in both Dries and Dag Wieers RPM repositories.

1. Cache::Cache (standard)

Implements Cache::MemoryCache, Cache::SharedMemoryCache (using IPC::ShareLite), Cache::FileCache and Cache::SizeAwareFileCache. This is the standard: various other modules often say that they implement Cache::Cache interface.

It shows a bit its age, so various improvements exists, including CHI
- unified cache interface (which can use Cache::Cache modules, and
also other caching backends like Cache::FastMmap in a unified way),
and Cache - the cache interface, which tries to improve Cache::Cache
but is not yet complete.

Here is some sample code for instantiating and using file system based cache (it uses Storable for serialization, IIRC).

  use Cache::FileCache;

  my $cache = new Cache::FileCache({default_expires_in => "15 minutes"});

  my $customer = $cache->get($name);

  if (not defined $customer)  {
    $customer = get_customer_from_db($name);
    $cache->set($name, $customer, "10 minutes");
  }

  return $customer;

Cache::Cache distribution can be found (at least on RPM based Linux distributions) in perl-Cache-Cache package (e.g. in Dries RPM repository).

Various other modules implements Cache::Cache interface, for example Cache::BerkeleyDB (compare with Cache::BDB).

2. CHI - Unified cache interface

CHI provides a unified caching API. The CHI interface is implemented by driver classes that support fetching, storing and clearing of data.

CHI is intended as an evolution (and successor) of DeWitt Clinton's Cache::Cache package, adhering to the basic Cache API but adding new features and addressing limitations in the Cache::Cache implementation.

Main goals of CHI were performance (minimizing method calls, serializing data only when necessary) and making the creation of new drivers as easy as possible.

The latter had lead to wrapping most popular caches available on CPAN in CHI interface (CHI handles serialization and expiration times) with CHI::Driver::CacheCache, CHI::Driver::FastMmap, CHI::Driver::Memcached. "Native" CHI drivers include 'File' (one file per entry), 'Memory' (per-process) and 'Multilevel' (two or more CHI caches, e.g. memcached bolstered by a local memory cache). 'DBI' and 'BerkeleyDB' are planned...

CHI provides expire_if [CODEREF] for additional check if cache expired, busy_lock [DURATION] to set expiration time to current time plus specified duration if value has expired to prevent "cache stampede", and expires_variance [FLOAT] to allow items to expire little earlier to prevent cache miss stampedes (favored over busy_lock). Even if gitweb wouldn't use CHI interface directly, those ideas are worth considering.

In addition to standard get() and set() methods it implements compute() method which combines get and set operations in a single call. It also has some methods to process multiple keys and/or values at once.

Here is some sample code for instantiating and using file system based cache.

    use CHI;

    # Choose a standard driver
    #
    my $cache = CHI->new(driver => 'File', 
                         cache_root => '/tmp/cache');

    # Basic cache operations
    #
    my $customer = $cache->get($name);
    if (!defined $customer) {
        $customer = get_customer_from_db($name);
        $cache->set($name, $customer, "10 minutes");
    }

    # or simply
    my $customer = $cache->compute($name, \&get_customer_from_db,
                                   "10 minutes");

3. Cache - the Cache interface 

The Cache modules are a total redesign and reimplementation of Cache::Cache and thus not directly compatible. Contrary to Cache::Cache get() and set() methods do not serialize complex data types; you have to freeze() and thaw() data explicitely, instead of set/get. You can get IO::Handle by which data can be read from, or written to cache, e.g. when using Cache::File. There is no concept of 'namespace' in the basic cache interface. Purging is done automatically in current implementation.

Currently only Cache::File (filesystem based implementation, could be done more efficiently, currently supports only LOCK_NFS locking) and Cache::Memory (per-process memory based implemetation; with namespaces) drivers are implemented.

In Cache modules one can select removal strategy for the cache.  By
default FIFO (First In First Out: remove oldest) and LRU (Least Recently
Used: remove stalest) strategies are available (when cache has size
limit?).

Cache modules provide callback interface: load_callback to be called anytime when a get() is issued for a data that does not exist in cache, and validate_callback (for example storing and checking timestamp or similar). This means that sample code for instantiating and using file system based cache can be written as below.

  use Cache::File;

  my $cache = Cache::File->new(cache_root => '/tmp/cacheroot');
  $cache->set_load_callback(\&get_customer_from_db);

  # calls get_customer_from_db() if needed
  my $customer = $cache->get($name);

The Cache classes can be used via the tie interface, as shown below. This allows the cache to be accessed via a hash. All the standard methods for accessing the hash are supported, with the exception of the 'keys' or 'each' call.

  tie %hash, 'Cache::File', { cache_root => $tempdir };

  $hash{'key'} = 'some data';
  $data = $hash{'key'};

The tie interface is especially useful with the load_callback to automatically populate the hash.

Even if gitweb wouldn't use Cache modules (perhaps because lack of matority, or/and the fact that they ar not in extras or trusted contrib packages repository) the idea of selectable removal strategy and the idea of callback interface are worth considering; perhaps even tie interface. Whether to serialize explicitely or not... that is also to be decided.

4. Other interesting caching packages

4.1. Cache::Adaptive for adaptive cache lifetime control

Cache::Adaptive is a cache engine with adaptive lifetime control. Cache lifetimes can be increased or decreased by any factor, e.g. load average, process time for building the cache entry, etc. Can use almost any Cache::Cache object as backend (the update algorithm needs reliable set() method, so Cache::SizeAwareFileCache cannot be used).

Cache::Adaptive::ByLoad is a subclass of Cache::Adaptive, which adjusts cache lifetime by two factors; the load average of the platform and the percentage of the total time spent by the builder.

Cache::Adaptive Introduces additional 
  access({ key => cache_key, builder => sub { ... } })
method, which returns cached entry if possible, or builds the entry by
calling the builder function, and optionally stores the build entry to
cache.  Compare with compute() method from CHI, or callback interfaces.

Worth examining (both interface and implementation) if/when implementing cache lifetime control based on load average, like kernel.org gitweb tries to do. J.H. (kernel.org) fork of gitweb uses longer lifetime under heavier load (within specified bounds).

4.2. Cache::Memcached and/or Cache::Swifty for caching using cache daemon

For larger installations, when there is needed caching not only for gitweb, it might be worth examining cache daemon solutions, like memcached (and Cache::Memcached, Cache::Memcached::Fast or CHI equivalent), a distributed memory cache daemon; or swifty (and Cache::Swifty), a very fast shared memory cache, in early alpha stages.

The Cache::Memcached api, besides set/get methods and administrative methods provide add() and replace() methods to set() conditionally, if value doesn't exists, or does exists in the cache.

Memcached was created to reduce load on high-trafic site with a hight database load that contains mostly read threads, so it might be not approproate for gitweb, where I/O load is of most concern, not CPU. The main advantage of memcached is its ability to scale out. Usually you can run memcached together with web serever or database server, as memcached is CPU lean and memory hungry, and web/database server the reverse: CPU hungry and usually memory lean. Note that gitweb (or rather git access to repositories in gitweb) is I/O hungry.

4.3. Cache::FastMmap (also example of callbacks),
     and caching benchmark mentioned there

Cache::FastMmap uses an mmap'ed file to act as a shared memory interprocess cache. It uses fcntl locking to ensure multiple processes can safely access the cache at the same time. It uses a basic LRU algorithm to keep the most used entries in the cache, plus (optionally) cache timeout.

Cache::FastMmap was created to be very fast.

The class also supports read-through, and write-back or write-through callbacks to access the real data if it's not in the cache. With those the code to deal with cache can be written simply as

  Cache::FastMmap->new(
    ...
    context => $RealDataSourceHandle,
    read_cb  => sub { $_[0]->get($_[1]) },
    write_cb => sub { $_[0]->set($_[1], $_[2]) },
  );

  ...

  my $value = $cache->get($key);

  $cache->set($key, $newvalue);

It also supports get_and_set() subroutine to atomically retrieve and set value of given key, and has methods dealing with multiple keys at once. There is also Cache::FastMmap::Tie module to use tie interface to Cache::FastMap. Even if gitweb wouldn't use this module, the callback based interface is worth considering to implement.

4.4. CGI::Cache to help cache output of time-intensive CGI scripts
     with minimal changes to CGI script code.

This module is intended to be used in CGI scripts that may benefit from caching; it is written in such a way that existing CGI code could get caching added with minimal changes to script. Here's a simple example:

  #!/usr/bin/perl

  use CGI;
  use CGI::Cache;

  # Set up cache
  CGI::Cache::setup();

  my $cgi = new CGI;

  # CGI::Vars requires CGI version 2.50 or better
  CGI::Cache::set_key($cgi->Vars);

  # This should short-circuit the rest of the loop if a cache value is
  # already there
  CGI::Cache::start() or exit;

  print $cgi->header, "\n";

  #...

  print <<EOF;
  This prints to STDOUT, which will be cached.
  If the next visit is within 24 hours, the cached STDOUT
  will be served instead of executing this 'print'.
  EOF

CGI::Cache module ties the output file descriptor (usually STDOUT) to an internal variable to which all output is saved. This trick (technique) is worth considering if we decide on caching final output in gitweb, or final output without HTTP headers.

5. Summary

If it is decided that gitweb would do caching of Perl structures, we would certainly use Storable, which should be installed as part of Perl installation on most systems. Perhaps gitweb could use Cache::Cache packages in general (and Cache::FileCache in particular), as it should fill "extras or trusted contrib" criterion, but I'd rather not add another dependency to gitweb, especially that not all installations need caching. It could be good solution for gitweb fork, and I guess kernel.org gitweb could use it.

If gitweb is to implement it's own solutions to not introduce extra dependencies, and it would cache Perl structures, implementing Cache::Cache get/set interface, with possible improvements of callback interface would be a good idea. For very large installations it would be good to check memcached solution (or multilevel cache, see CHI).

If gitweb is to use caching of output, or output without HTTP headers, either using CGI::Cache or using its technique would be a good idea.

Gitweb caching is meant to reduce load (mainly I/O load according to some mails send on this mailing list by J.H., kernel.org gitweb admin, and Pasky, repo.or.cz gitweb admin). I think it would be good to try and check, benchmarking if possible, different solutions to "thundering horde" aka "cache stampede" problem, and to adaptive cache lifetime control (see Cache::Adaptive).

Thoughts? Comments?

%% In the next part I'd like to have thoughts and ideas for gitweb caching from J.H. and Petr 'Pasky' Baudis...

References: =========== [1] http://search.cpan.org [2] http://code.google.com/p/perl-cache [3] http://www.danga.com/memcached/ [4] http://cpan.robm.fastmail.fm/cache_perf.html