The archivist’s web crawler: WARC output, dashboard for all crawls, dynamic ignore patterns

grab-site

grab-site is an easy preconfigured web crawler designed for backing up websites. Give grab-site a URL and it will recursively crawl the site and write WARC files. Internally, grab-site uses a fork of wpull for crawling.

grab-site gives you

• a dashboard with all of your crawls, showing which URLs are being grabbed, how many URLs are left in the queue, and more.

• the ability to add ignore patterns when the crawl is already running. This allows you to skip the crawling of junk URLs that would otherwise prevent your crawl from ever finishing. See below.

• an extensively tested default ignore set (global) as well as additional (optional) ignore sets for forums, reddit, etc.

• duplicate page detection: links are not followed on pages whose content duplicates an already-seen page.

The URL queue is kept on disk instead of in memory. If you’re really lucky, grab-site will manage to crawl a site with ~10M pages.

Note: if you have any problems whatsoever installing or getting grab-site to run, please file an issue - thank you!

Contents

• Install on Ubuntu 16.04, 18.04, Debian 9 (stretch), Debian 10 (buster)
• Install on NixOS
• Install on another distribution lacking Python 3.7.x
• Install on macOS
• Install on Windows 10 (experimental)
• Upgrade an existing install
• Usage
  • grab-site options, ordered by importance
  • Warnings
  • Tips for specific websites
• Changing ignores during the crawl
• Inspecting the URL queue
• Preventing a crawl from queuing any more URLs
• Stopping a crawl
• Advanced gs-server options
• Viewing the content in your WARC archives
• Inspecting WARC files in the terminal
• Automatically pausing grab-site processes when free disk is low
• Thanks
• Help

Install on Ubuntu 16.04, 18.04, Debian 9 (stretch), Debian 10 (buster)

On Debian, use su to become root if sudo is not configured to give you access.

sudo apt-get update
sudo apt-get install --no-install-recommends \
    git build-essential libssl-dev zlib1g-dev libbz2-dev libreadline-dev \
    libsqlite3-dev libffi-dev libxml2-dev libxslt1-dev libre2-dev pkg-config

If you see Unable to locate package, run the two commands again.

As a non-root user:

wget https://raw.githubusercontent.com/pyenv/pyenv-installer/master/bin/pyenv-installer
chmod +x pyenv-installer
./pyenv-installer
~/.pyenv/bin/pyenv install 3.7.5
~/.pyenv/versions/3.7.5/bin/python -m venv ~/gs-venv
~/gs-venv/bin/pip install --no-binary lxml --upgrade git+https://github.com/ArchiveTeam/grab-site

--no-binary lxml is necessary for the html5-parser build.

Add this to your ~/.bashrc or ~/.zshrc and then restart your shell (e.g. by opening a new terminal tab/window):

PATH="$PATH:$HOME/gs-venv/bin"

Install on NixOS

As a non-root user:

nix-env -f https://github.com/NixOS/nixpkgs/archive/master.tar.gz -iA grab-site

Install on another distribution lacking Python 3.7.x

grab-site and its dependencies are available in nixpkgs, which can be used on any Linux distribution.

1. As root:

   Where USER is your non-root username:

   mkdir /nix
   chown USER:USER /nix
   

2. As the non-root user, install Nix: https://nixos.org/nix/download.html

3. As the non-root user:

   nix-env -f https://github.com/NixOS/nixpkgs/archive/master.tar.gz -iA grab-site
   

And then restart your shell (e.g. by opening a new terminal tab/window).

Install on macOS

Please note macOS is currently affected by issue #157, help with a fix would be appreciated.

On OS X 10.10 - macOS 10.13:

1. Run locale in your terminal. If the output includes “UTF-8”, you are all set. If it does not, your terminal is misconfigured and grab-site will fail to start. This can be corrected with:

   • Terminal.app: Preferences… -> Profiles -> Advanced -> check Set locale environment variables on startup

   • iTerm2: Preferences… -> Profiles -> Terminal -> Environment -> check Set locale variables automatically

Using Homebrew

2. Install Homebrew using the install step on https://brew.sh/

3. Run:

   brew update
   brew install python libxslt re2 pkg-config
   python3 -m venv ~/gs-venv
   PKG_CONFIG_PATH="/usr/local/opt/libxml2/lib/pkgconfig" ~/gs-venv/bin/pip install --no-binary lxml --upgrade git+https://github.com/ArchiveTeam/grab-site
   

4. Add this to your ~/.bash_profile (which may not exist yet) and then restart your shell (e.g. by opening a new terminal tab/window):

   PATH="$PATH:$HOME/gs-venv/bin"
   

   ``

Using Nix

As an alternative to the Homebrew install, if you prefer Nix.

2. Install Nix: https://nixos.org/nix/download.html

3. Run:

   nix-env -f https://github.com/NixOS/nixpkgs/archive/master.tar.gz -iA grab-site
   

And then restart your shell (e.g. by opening a new terminal tab/window).

Install on Windows 10 (experimental)

On Windows 10 Fall Creators Update (1703) or newer:

1. Start menu -> search “feature” -> Turn Windows features on or off

2. Scroll down, check “Windows Subsystem for Linux” and click OK.

3. Wait for install and click “Restart now”

4. Start menu -> Store

5. Search for “Ubuntu” in the store and install Ubuntu (publisher: Canonical Group Limited).

6. Start menu -> Ubuntu

7. Wait for install and create a user when prompted.

8. Follow the Ubuntu 16.04, 18.04, Debian 9 (stretch), Debian 10 (buster) steps.

Upgrade an existing install

To update grab-site, simply run the ~/gs-venv/bin/pip install ... or nix-env ... command used to install it originally (see above).

After upgrading, stop gs-server with kill or ctrl-c, then start it again. Existing grab-site crawls will automatically reconnect to the new server.

Usage

First, start the dashboard with:

gs-server

and point your browser to http://127.0.0.1:29000/

Note: gs-server listens on all interfaces by default, so you can reach the dashboard by a non-localhost IP as well, e.g. a LAN or WAN IP. (Sub-note: no code execution capabilities are exposed on any interface.)

Then, start as many crawls as you want with:

grab-site 'URL'

Do this inside tmux unless they’re very short crawls. Note that tmux 2.1 is broken and will lock up frequently. Ubuntu 16.04 users probably need to remove tmux 2.1 and install tmux 1.8 from Ubuntu 14.04. If you are unable to downgrade tmux, detaching immediately after starting the crawl may be enough to avoid the problem.

grab-site outputs WARCs, logs, and control files to a new subdirectory in the directory from which you launched grab-site, referred to here as “DIR”. (Use ls -lrt to find it.)

You can pass multiple URL arguments to include them in the same crawl, whether they are on the same domain or different domains entirely.

warcprox users: warcprox breaks the dashboard’s WebSocket; please make your browser skip the proxy for whichever host/IP you’re using to reach the dashboard.

grab-site options, ordered by importance

Options can come before or after the URL.

• --1: grab just URL and its page requisites, without recursing.

• --igsets=IGSET1,IGSET2: use ignore sets IGSET1 and IGSET2.

  Ignore sets are used to avoid requesting junk URLs using a pre-made set of regular expressions. See the full list of available ignore sets.

  The global ignore set is implied and always enabled.

  The ignore sets can be changed during the crawl by editing the DIR/igsets file.

• --no-offsite-links: avoid following links to a depth of 1 on other domains.

  grab-site always grabs page requisites (e.g. inline images and stylesheets), even if they are on other domains. By default, grab-site also grabs linked pages to a depth of 1 on other domains. To turn off this behavior, use --no-offsite-links.

  Using --no-offsite-links may prevent all kinds of useful images, video, audio, downloads, etc from being grabbed, because these are often hosted on a CDN or subdomain, and thus would otherwise not be included in the recursive crawl.

• -i / --input-file: Load list of URLs-to-grab from a local file or from a URL; like wget -i. File must be a newline-delimited list of URLs. Combine with --1 to avoid a recursive crawl on each URL.

• --igon: Print all URLs being ignored to the terminal and dashboard. Can be changed during the crawl by touching or rming the DIR/igoff file. This is slower because it needs to find the specific regexp to blame.

• --no-video: Skip the download of videos by both mime type and file extension. Skipped videos are logged to DIR/skipped_videos. Can be changed during the crawl by touching or rming the DIR/video file.

• --no-sitemaps: don’t queue URLs from sitemap.xml at the root of the site.

• --max-content-length=N: Skip the download of any response that claims a Content-Length larger than N. (default: -1, don’t skip anything). Skipped URLs are logged to DIR/skipped_max_content_length. Can be changed during the crawl by editing the DIR/max_content_length file.

• --no-dupespotter: Disable dupespotter, a plugin that skips the extraction of links from pages that look like duplicates of earlier pages. Disable this for sites that are directory listings, because they frequently trigger false positives.

• --concurrency=N: Use N connections to fetch in parallel (default: 2). Can be changed during the crawl by editing the DIR/concurrency file.

• --delay=N: Wait N milliseconds (default: 0) between requests on each concurrent fetcher. Can be a range like X-Y to use a random delay between X and Y. Can be changed during the crawl by editing the DIR/delay file.

• --import-ignores: Copy this file to to DIR/ignores before the crawl begins.

• --warc-max-size=BYTES: Try to limit each WARC file to around BYTES bytes before rolling over to a new WARC file (default: 5368709120, which is 5GiB). Note that the resulting WARC files may be drastically larger if there are very large responses.

• --level=N: recurse N levels instead of inf levels.

• --page-requisites-level=N: recurse page requisites N levels instead of 5 levels.

• --ua=STRING: Send User-Agent: STRING instead of pretending to be Firefox on Windows.

• --id=ID: Use id ID for the crawl instead of a random 128-bit id. This must be unique for every crawl.

• --dir=DIR: Put control files, temporary files, and unfinished WARCs in DIR (default: a directory name based on the URL, date, and first 8 characters of the id).

• --finished-warc-dir=FINISHED_WARC_DIR: absolute path to a directory into which finished .warc.gz and .cdx files will be moved.

• --permanent-error-status-codes=STATUS_CODES: A comma-separated list of HTTP status codes to treat as a permanent error and therefore not retry (default: 401,403,404,405,410). Other error responses tried another 2 times for a total of 3 tries (customizable with --wpull-args=--tries=N). Note that, unlike wget, wpull puts retries at the end of the queue.

• --wpull-args=ARGS: String containing additional arguments to pass to wpull; see wpull --help. ARGS is split with shlex.split and individual arguments can contain spaces if quoted, e.g. --wpull-args="--youtube-dl \"--youtube-dl-exe=/My Documents/youtube-dl\""

  Also useful: --wpull-args=--no-skip-getaddrinfo to respect /etc/hosts entries.

• --which-wpull-args-partial: Print a partial list of wpull arguments that would be used and exit. Excludes grab-site-specific features, and removes DIR/ from paths. Useful for reporting bugs on wpull without grab-site involvement.

• --which-wpull-command: Populate DIR/ but don’t start wpull; instead print the command that would have been used to start wpull with all of the grab-site functionality.

• --debug: print a lot of debug information.

• --help: print help text.

Warnings

If you pay no attention to your crawls, a crawl may head down some infinite bot trap and stay there forever. The site owner may eventually notice high CPU use or log activity, then IP-ban you.

grab-site does not respect robots.txt files, because they frequently whitelist only approved robots, hide pages embarrassing to the site owner, or block image or stylesheet resources needed for proper archival. See also. Because of this, very rarely you might run into a robot honeypot and receive an abuse@ complaint. Your host may require a prompt response to such a complaint for your server to stay online. Therefore, we recommend against crawling the web from a server that hosts your critical infrastructure.

Don’t run grab-site on GCE (Google Compute Engine); as happened to me, your entire API project may get nuked after a few days of crawling the web, with no recourse. Good alternatives include OVH (OVH, So You Start, Kimsufi), and online.net’s dedicated and Scaleway offerings.

Tips for specific websites

Website requiring login / cookies

Log in to the website in Chrome or Firefox. Use the cookies.txt extension for Chrome or for Firefox extension to copy Netscape-format cookies. Paste the cookies data into a new file. Start grab-site with --wpull-args=--load-cookies=ABSOLUTE_PATH_TO_COOKIES_FILE.

Static websites; WordPress blogs; Discourse forums

The defaults usually work fine.

Blogger / blogspot.com blogs

The defaults work fine except for blogs with a JavaScript-only Dynamic Views theme.

Some blogspot.com blogs use “Dynamic Views” themes that require JavaScript and serve absolutely no HTML content. In rare cases, you can get JavaScript-free pages by appending ?m=1 (example). Otherwise, you can archive parts of these blogs through Google Cache instead (example) or by using https://archive.is/ instead of grab-site.

Tumblr blogs

Either don’t crawl from Europe (because tumblr redirects to a GDPR /privacy/consent page), or add Googlebot to the user agent:

--ua "Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:64.0) Gecko/20100101 Firefox/70.0 but not really nor Googlebot/2.1"

Use --igsets=singletumblr to avoid crawling the homepages of other tumblr blogs.

If you don’t care about who liked or reblogged a post, add \?from_c= to the crawl’s ignores.

Some tumblr blogs appear to require JavaScript, but they are actually just hiding the page content with CSS. You are still likely to get a complete crawl. (See the links in the page source for https://X.tumblr.com/archive).

Subreddits

Use --igsets=reddit and add a / at the end of the URL to avoid crawling all subreddits.

When crawling a subreddit, you must get the casing of the subreddit right for the recursive crawl to work. For example,

grab-site https://www.reddit.com/r/Oculus/ --igsets=reddit

will crawl only a few pages instead of the entire subreddit. The correct casing is:

grab-site https://www.reddit.com/r/oculus/ --igsets=reddit

You can hover over the “Hot”/“New”/… links at the top of the page to see the correct casing.

Directory listings (“Index of …")

Use --no-dupespotter to avoid triggering false positives on the duplicate page detector. Without it, the crawl may miss large parts of the directory tree.

Very large websites

Use --no-offsite-links to stay on the main website and avoid crawling linked pages on other domains.

Websites that are likely to ban you for crawling fast

Use --concurrency=1 --delay=500-1500.

MediaWiki sites with English language

Use --igsets=mediawiki. Note that this ignore set ignores old page revisions.

MediaWiki sites with non-English language

You will probably have to add ignores with translated Special:* URLs based on ignore_sets/mediawiki.

Forums that aren’t Discourse

Forums require more manual intervention with ignore patterns. --igsets=forums is often useful for non-SMF forums, but you will have to add other ignore patterns, including one to ignore individual-forum-post pages if there are too many posts to crawl. (Generally, crawling the thread pages is enough.)

GitHub issues / pull requests

Find the highest issue number from an issues page (example) and use:

grab-site --1 https://github.com/rust-lang/rust/issues/{1..30000}

This relies on your shell to expand the argument to thousands of arguments. If there are too many arguments, you may have to write the URLs to a file and use grab-site -i instead:

for i in {1..30000}; do echo https://github.com/rust-lang/rust/issues/$i >> .urls; done
grab-site --1 -i .urls

Websites whose domains have just expired but are still up at the webhost

Use a DNS history service to find the old IP address (the DNS “A” record) for the domain. Add a line to your /etc/hosts to point the domain to the old IP. Start a crawl with --wpull-args=--no-skip-getaddrinfo to make wpull use /etc/hosts.

twitter.com/user

Use snscrape to get a list of tweets for a user. Redirect snscrape’s output to a list of URLs with > urls and pass this file to grab-site --1 -i urls.

Alternatively, use webrecorder.io instead of grab-site. It has an autoscroll feature and you can download the WARCs.

Keep in mind that scrolling twitter.com/user returns a maximum of 3200 tweets, while a from:user query can return more.

Changing ignores during the crawl

While the crawl is running, you can edit DIR/ignores and DIR/igsets; the changes will be applied within a few seconds.

DIR/igsets is a comma-separated list of ignore sets to use.

DIR/ignores is a newline-separated list of Python 3 regular expressions to use in addition to the ignore sets.

You can rm DIR/igoff to display all URLs that are being filtered out by the ignores, and touch DIR/igoff to turn it back off.

Note that ignores will not apply to any of the crawl’s start URLs.

Inspecting the URL queue

Inspecting the URL queue is usually not necessary, but may be helpful for adding ignores before grab-site crawls a large number of junk URLs.

To dump the queue, run:

gs-dump-urls DIR/wpull.db todo

Four other statuses can be used besides todo: done, error, in_progress, and skipped.

You may want to pipe the output to sort and less:

gs-dump-urls DIR/wpull.db todo | sort | less -S

Preventing a crawl from queuing any more URLs

rm DIR/scrape. Responses will no longer be scraped for URLs. Scraping cannot be re-enabled for a crawl.

Stopping a crawl

You can touch DIR/stop or press ctrl-c, which will do the same. You will have to wait for the current downloads to finish.

Advanced gs-server options

These environmental variables control what gs-server listens on:

• GRAB_SITE_INTERFACE (default 0.0.0.0)
• GRAB_SITE_PORT (default 29000)

These environmental variables control which server each grab-site process connects to:

• GRAB_SITE_HOST (default 127.0.0.1)
• GRAB_SITE_PORT (default 29000)

Viewing the content in your WARC archives

You can use ikreymer/webarchiveplayer to view the content inside your WARC archives. It requires Python 2, so install it with pip instead of pip3:

sudo apt-get install --no-install-recommends git build-essential python-dev python-pip
pip install --user git+https://github.com/ikreymer/webarchiveplayer

And use it with:

~/.local/bin/webarchiveplayer <path to WARC>

then point your browser to http://127.0.0.1:8090/

Inspecting WARC files in the terminal

zless is a wrapper over less that can be used to view raw WARC content:

zless DIR/FILE.warc.gz

zless -S will turn off line wrapping.

Note that grab-site requests uncompressed HTTP responses to avoid double-compression in .warc.gz files and to make zless output more useful. However, some servers will send compressed responses anyway.

Automatically pausing grab-site processes when free disk is low

If you automatically upload and remove finished .warc.gz files, you can still run into a situation where grab-site processes fill up your disk faster than your uploader process can handle. To prevent this situation, you can customize and run this script, which will pause and resume grab-site processes as your free disk space crosses a threshold value.

Thanks

grab-site is made possible only because of wpull, written by Christopher Foo who spent a year making something much better than wget. ArchiveTeam’s most pressing issue with wget at the time was that it kept the entire URL queue in memory instead of on disk. wpull has many other advantages over wget, including better link extraction and Python hooks.

Thanks to David Yip, who created ArchiveBot. The wpull hooks in ArchiveBot served as the basis for grab-site. The original ArchiveBot dashboard inspired the newer dashboard now used in both projects.

Thanks to Falcon Darkstar Momot for the many wpull 2.x fixes that were rolled into ArchiveTeam/wpull.

Thanks to JustAnotherArchivist for investigating my wpull issues.

Help

grab-site bugs and questions are welcome in grab-site/issues.

Terminal output in your bug report should be surrounded by triple backquotes, like this:

Please report security bugs as regular bugs.