I’m getting lots of 504 Timeout errors, please help!¶
HTTP 504 error means a request to Splash took more than timeout seconds to complete (30s by default) - Splash aborts script execution after the timeout. To override the timeout value pass ‘timeout’ argument to the Splash endpoint you’re using.
Note that the maximum allowed
timeout value is limited by the maximum
timeout setting, which is by default 60 seconds. In other words,
by default you can’t pass
?timeout=300 to run a long script - an
error will be returned.
Maximum allowed timeout can be increased by passing
option to Splash server on startup:
$ python -m splash.server --max-timeout 3600
For Docker the command would be something like this (see Passing Custom Options):
$ docker run -it -p 8050:8050 scrapinghub/splash --max-timeout 3600
The next question is why a request can need 10 minutes to render. There are 3 common reasons:
1. Slow website¶
A website can be really slow, or it can try to get some remote resources which are really slow.
There is no way around increasing timeouts and reducing request rate if the website itself is slow. However, often the problem lays in unreliable remote resources like third-party trackers or advertisments. By default Splash waits for all remote resources to load, but in most cases it is better not to wait for them forever.
To abort resource loading after a timeout and give the whole page a chance to
render use resource timeouts. For render.*** endpoints use
for execute use either splash.resource_timeout or
request:set_timeout (see splash:on_request).
It is a good practive to always set resource_timeout; something similar to
resource_timeout=20 often works well.
2. Splash Lua script does too many things¶
When a script fetches many pages or uses large delays then timeouts
are inevitable. Sometimes you have to run such scripts; in this case increase
--max-timeout Splash option and use larger timeout
But before increasing the timeouts consider splitting your script into smaller steps and sending them to Splash individually. For example, if you need to fetch 100 websites, don’t write a Splash Lua script which takes a list of 100 URLs and fetches them - write a Splash Lua script that takes 1 URL and fetches it, and send 100 requests to Splash. This approach has a number of benefits: it makes scripts more simple and robust and enables parallel processing.
3. Splash instance is overloaded¶
When Splash is overloaded it may start producing 504 errors.
Splash renders requests in parallel, but it doesn’t render them all
at the same time - concurrency is limited to a value set at startup
--slots option. When all slots are used a request is put into
a queue. The thing is that a timeout starts to tick once Splash receives
a request, not when Splash starts to render it. If a request stays in an
internal queue for a long time it can timeout even if a website is fast
and splash is capable of rendering the website.
To increase rendering speed and fix an issue with a queue it is recommended to start several Splash instances and use a load balancer capable of maintaining its own request queue. HAProxy has all necessary features; check an example config here. A shared request queue in a load balancer also helps with reliability: you won’t be loosing requests if a Splash instance needs to be restarted.
How to run Splash in production?¶
Don’t forget to use resource timeous in your client code (see 1. Slow website). It also makes sense to retry a couple of times if Splash returns 5xx error response.
If you want to create your own production setup, here is a small non-exhaustive checklist:
- Splash should be daemonized and started on boot;
- in case of failures or segfaults Splash must be restarted;
- memory usage should be limited;
- several Splash instances should be started to use all CPU cores and/or multiple servers;
- requests queue should be moved to the load balancer to make rendering more robust (see 3. Splash instance is overloaded).
Of course, it is also good to setup monitoring, configuration management, etc. - all the usual stuff.
To daemonize Splash, start it on boot and restart on failures
one can use Docker: since Docker 1.2 there are
-d options which can be used together. Another way to do that is
to use standard tools like upstart, systemd
--restart option won’t work without
Splash uses an unbound in-memory cache and so it will eventually consume
all RAM. A workaround is to restart the process when it uses too much memory;
there is Splash
--maxrss option for that. You can also add Docker
--memory option to the mix.
In production it is a good idea to pin Splash version - instead of
scrapinghub/splash it is usually better to use something like
A command for starting a long-running Splash server which uses up to 4GB RAM and daemonizes & restarts itself could look like this:
$ docker run -d -p 8050:8050 --memory=4.5G --restart=always scrapinghub/splash:2.0 --maxrss 4000
How do I disable Private mode?¶
With Splash>=2.0, you can disable Private mode (which is “on” by default). There are two ways to go about it:
at startup, with the
--disable-private-modeargument, e.g., if you’re using Docker:
$ sudo docker run -it -p 5023:5023 -p 8050:8050 -p 8051:8051 scrapinghub/splash --disable-private-mode
at runtime when using the
/executeendpoint and setting splash.private_mode_enabled attribute to
Note that if you disable private mode then browsing data such as cookies or items kept in localStorage may persist between requests. If you’re using Splash in a shared environment it could mean your cookies or local storage items can be accessed by other clients, or that you can occasionally access other client’s cookies.
Check this github issue for the motivation.
render.html result looks broken in a browser¶
When you check
in a browser it is likely stylesheets & other resources won’t
load properly. It happens when resource URLs are relative - the browser
will resolve them as relative to
http://<splash-server>:8050/render.html?url=<url>, not to
This is not a Splash bug, it is a standard browser behaviour.
If you just want to check how the page looks like after rendering use render.png or render.jpeg endpoints. If screenshot is not an option and you want to display html with images, etc. using a browser then you may post-process the HTML and add an appropriate <base> HTML tag to the page.
But by passing baseurl you’re instructing Splash to use it, not your browser. It doesn’t change relative links to absolute in DOM, it makes Splash to treat them as relative to baseurl when rendering.
Changing links to absolute in DOM tree is not what browsers do when
base url is applied - e.g. if you check href attribute using JS code
it will still contain relative value even if
<base> tag is used.
render.html returns DOM snapshot, so the links are not changed.
When you load render.html result in a browser it is your browser who resolves relative links, not Splash, so they are resolved incorrectly.