This chapter is a “deep dive” into Bokeh server’s internals. It assumes you’re already familiar with the information on Bokeh server in the :ref`userguide`.
You might want to read this if you are:
- trying to work on the Bokeh codebase
- writing your own custom server process to use rather than
A custom server process can add additional routes (web pages or REST endpoints) using Tornado’s web framework.
If an application developer uses
bokeh serve they typically should not need
to import from
bokeh.server at all. An application developer would only use
Server class if it is doing something specialized, such as a custom
or embedded server process.
Applications, Sessions, and Connections¶
Each server contains one or more applications; you can think of an application
as a session template, or a factory for sessions. Sessions have a 1-1
relationship with instances of
bokeh.document.Document: each session has a
document instance. When a browser connects to the server, it gets a new
session; the application fills in the session’s document with whatever plots,
widgets, or other content it desires. The application can also set up
callbacks, to run periodically or to run when the document changes.
Applications are represented by the
Application class. This class is
little more than a list of
Handler instances. Handlers can be created
in lots of ways; from JSON files, from Python functions, from Python files,
and perhaps many more ways in the future.
Around each application, the server creates an
primary role is to hold the set of sessions for the application.
Sessions are represented by class
Each application has a route (called an
app_path in the client
API), and each session has an ID. The combination of the two
Document instance (the server looks up the
application by path, and then looks up the session by ID).
Each session has 0-N connections, represented by the
class. Connections are websocket connections. In general, sessions last as
long as they have connections, though they only expire after a timeout (to
allow for page reloads and the like).
Applications and application handlers cannot access the
ApplicationContext directly; they have a much more
limited interface defined in two pieces,
ServerContext presents a limited interface to some
presents a limited interface to some aspects of
implementations of these interfaces are
Summarizing the object graph:
- has N
- has 1
Applicationcapable of creating new sessions
- has 1 path used to identify it in URLs
- has 1
ServerContextrepresenting the aspects of the server visible to application code
- has N
- has 1 session ID which is a string naming the session
- has 1
Documentrepresenting the session state
- has N
ServerConnectionrepresenting websockets attached to the session
- has 1
SessionContextrepresenting the aspects of the session visible to application code
IOLoop and Async Code¶
To work on the server, you’ll need an understanding of Tornado’s
IOLoop and the
The Tornado documentation will be the best resource, but here are some quick things to know:
- the Bokeh server is single-threaded, so it’s important not to write “blocking” code, meaning code that uses up the single thread while it waits for IO or performs a long computation. If you do this, you’ll rapidly increase the latency seen by users of your application. For example, if you block for 100ms every time someone moves a slider, and ten users are doing this at once, users could easily see 10*100ms=1s of lag... with only ten users.
- in Tornado, nonblocking code is modeled with functions or
methods that return an instance of the
Futureclass. You may have seen the
@gen.coroutinedecorator, which transforms the decorated method into a method which returns a
- when no code is running, Tornado waits in its
IOLoop(sometimes called a “main loop” or “event loop”), which means it’s waiting for something to happen. When something happens,
IOLoopexecutes any callbacks that were interested in that event.
Applications and the
We don’t want applications to touch the Tornado
directly to add callbacks, because when a session expires or an
application is reloaded, we need the ability to remove all
callbacks belonging to a session or application.
To enable this, applications should only add callbacks using the
ServerContext. Methods on those
classes allow applications to
intercept these callback additions and are able to remove them
when we unload an application or destroy a session.
If you look at the
Application class, there are two ways the
server can call into it.
modify_document()method which does just what it says: it passes in the session’s
Documentand allows the application to modify it (perhaps adding some plots and widgets).
- a set of “hooks”
The “hooks” are called “lifecycle hooks” since they happen at defined points in the lifetime of an application and a session.
Here are the steps in the lifecycle:
- When the server process starts up, it calls
on_server_loaded()on each application.
- When a client connects with a previously-unused session ID, the
server creates a
on_session_created()with an empty
modify_document()to initialize the
on_session_created()can also initialize part of the
Documentif it likes.
- When there are no connections to a session, it will eventually
time out and
on_session_destroyed()will be called.
- If the server process shuts down cleanly, it will call
on_server_unloaded()on each application. This is probably rare in production: it’s typical for server processes to be killed by a signal.
on_server_unloaded()may be more useful during development so that apps can be reloaded without leaking resources.
These hooks can add periodic or one-shot callbacks to the
ServerContext. These callbacks may be asynchronous (using
Tornado’s async IO facilities), and are able to update all live
Critical consideration when using ``on_server_loaded()``: Process-global is NOT the same as cluster-global. If you scale a Bokeh application, you’ll want a separate process for each CPU core, roughly. Processes in a cluster may not even be on the same machine. A server process can never assume that it knows about “all sessions that exist,” only “all sessions hosted in this process.”
The session object handles most interaction between the client and the server.
The trickiest aspect of
ServerSession may be locking. In general, we
want one callback or one websocket request to be processed at a time; we
don’t want to interleave them, because it would be difficult to implement
callbacks and request handlers if they had to worry about interleaving.
ServerSession does one thing at a time, controlled by
ServerSession._lock, which is a Tornado lock.
If you’re familiar with locking and threads, the situation here is conceptually
identical; but race conditions can only happen at “yield points” (when we
return to the
IOLoop) rather than at any point, and the lock is a Tornado
lock rather than a thread lock.
The rule is: to touch ServerSession.document code must hold ServerSession._lock.
For callbacks added through the
Document API, we automatically
acquire the lock on the callback’s behalf before we execute the
callback, and release it afterward.
For callbacks added through the
ServerContext API, they can only obtain
a reference to the session document using the method
with_locked_document() executes a function with
the document lock held, passing the document to that function. The lock is
held while the function runs (even if the function is asynchronous! if the
function returns a
Future, the lock is held until the
It is very easy to modify the server code in such a way that you’re touching the document without holding the lock. If you do this, things will break in subtle and painful-to-debug ways. When you touch the session document, triple-check that the lock is held.
For background on session IDs, check out the
bokeh serve documentation on
We rely on session IDs being cryptographically random and difficult to guess. If an attacker knows someone’s session ID, they can eavesdrop on or modify the session. If you’re writing a larger web app with a Bokeh app embedded inside, this may affect how you design your larger app.
When hacking on the server, for the most part session IDs are opaque strings and after initially validating the ID, it doesn’t matter to the server code what the ID is.
To avoid resource exhaustion, the server times out unused sessions. You can
find the code for this in
The server has a websocket connection open to each client (each browser tab,
in typical usage). The primary role of the websocket is to keep the session’s
Document in sync between the client and the server.
There are two client implementations in the Bokeh codebase; one is a Python
Client and server sessions are mostly symmetrical; on both sides, we are
receiving change notifications from the other side’s
Document, and sending
notification of changes made on our side. In this way, the two
are kept in sync.
The Python implementation of the websocket protocol can be found in
bokeh.server.protocol, though both the client side and the server side
Websockets already implement “frames” for us, and they guarantee frames will arrive in the same order they were sent. Frames are strings or byte arrays (or special internal frame types, such as pings). A websocket looks like a two sequences of frames, one sequence in each direction (“full duplex”).
On top of websocket frames, we implement our own
Message concept. A Bokeh
Message spans multiple websocket frames. It always contains a header frame,
metadata frame, and content frame. These three frames each contain a JSON
string. The code permits these three frames to be followed by binary data
frames, but currently in Bokeh binary data frames are not used.
The header frame indicates the message type and gives messages an ID. Message IDs are used to match replies with requests (the reply contains a field saying “I am the reply to the request with ID xyz”).
The metadata frame has nothing in it for now, but could be used for debugging data or another purpose in the future.
The content frame has the “body” of the message.
There aren’t many messages right now. A quick overview:
ACKis used for an initial handshake when setting up the connection
OKis a generic reply when a request doesn’t require any more specific reply
ERRORis a generic error reply when something goes wrong
SERVER-INFO-REPLYare a request-reply pair where the reply contains information about the server, such as its Bokeh version
PULL-DOC-REQasks to get the entire contents of the session’s
Documentas JSON, and
PULL-DOC-REPLYis the reply containing said JSON.
PUSH-DOCsends the entire contents of the session’s
Documentas JSON, and the other side should replace its document with these new contents.
PATCH-DOCsends changes to the session’s document to the other side
Typically, when opening a connection one side will pull or push
the entire document; after the initial pull or push, the two sides
stay in sync using
Some Current Protocol Caveats¶
- In the current protocol, conflicts where both sides change the
same thing at the same time are not handled (the two sides can
end up out-of-sync if this happens, because the two
PATCH-DOCare in flight at the same time). It’s easy to devise a scheme to detect this situation, but it’s less clear what to do when it’s detected, so right now we don’t detect it and do nothing. In most cases, applications should avoid this situation because even if we could make sense of it and handle it somehow, it would probably be inefficient for the two sides of the app to “fight” over the same value. (If real-world applications trip on this issue, we will have to figure out what they’re trying to do and devise a solution.)
- At the moment, we are not smart about patching collections; if
Modelproperty that’s a giant dictionary, we’ll send the whole giant dictionary whenever any entry in it changes.
- At the moment, we do not optimize binary data by sending it over binary websocket frames.
The server only supports a few HTTP routes; you can find them in
/static/serves Bokeh’s JS and CSS resources
/app_path/serves a page that displays a new session
/app_path/wsis the websocket connection URL
Bokeh server isn’t intended to be a general-purpose web framework. You can
however pass new endpoints to
Server using the
and the Tornado APIs.
In general whenever a model property is modified, the new value is
first validated, and the
Document is notified of the change. Just
as models may have
on_change callbacks, so can a
Document. When a
Document is notified of a change to one of
its models it will generate the appropriate event (usually a
ModelChangedEvent) and trigger the
passing them this new event. Sessions are one such callback, which
will turn the event into a patch that can be sent across the web
socket connection. When a message is received by the client or server
session it will extract the patch and apply it directly to the
In order to avoid events bouncing back and forth between client and
server (as each patch would generate new events, which would in turn
be sent back), the session informs the
Document that it was
responsible for generating the patch and any subsequent events that
are generated. In this way, when a
Session is notified of a change
to the document it can check whether the
event.setter is identical
with itself and therefore skip processing the event.
In general all the concepts above are agnostic as to how precisely the
models and change events are encoded and decoded. Each model and its
properties are responsible for converting their values to a JSON-like
format, which can be sent across the Websocket connection. One
difficulty here is that one model can reference other models, often in
highly interconnected and even circular ways. Therefore during the
conversion to a JSON-like format all references by one model to other
models are replaced with ID references. Additionally models and
properties can define special serialization behaviors, one such
example is the
ColumnData property on a
which will convert NumPy arrays to a base64 encoded representation,
which is significantly more efficient than sending numeric arrays in a
string based format. The
serializable_value method applies this encoding and the from_json
method will convert the data back. Equivalently the JS-based
ColumnDataSource knows how to interpret the base64 encoded data
attributes_as_json methods also knows how to encode the data. In
this way models can implement optimized serialization formats.
To test client-server functionality, use the utilities in
ManagedServerLoop, you can start up a server instance
server.io_loop with a client and you can
test any aspect of the server. Check out the existing tests for
lots of examples. Anytime you add a new websocket message or http
endpoint, be sure to add tests!