Running a Bokeh Server

Purpose

The architecture of Bokeh is such that high-level “model objects” (representing things like plots, ranges, axes, glyphs, etc.) are created in Python, and then converted to a JSON format that is consumed by the client library, BokehJS. (See Defining Key Concepts for a more detailed discussion.) By itself, this flexible and decoupled design offers advantages, for instance it is easy to have other languages (R, Scala, Lua, ...) drive the exact same Bokeh plots and visualizations in the browser.

However, if it were possible to keep the “model objects” in python and in the browser in sync with one another, then more additional and powerful possibilities immediately open up:

  • respond to UI and tool events generated in a browser with computations or queries using the full power of python
  • automatically push updates the UI (i.e. widgets or plots), in a browser
  • use periodic, timeout, and asynchronous callbacks drive streaming updates

This capability to synchronize between python and the browser is the main purpose of the Bokeh Server.

Use Case Scenarios

Now that we know what the Bokeh server is for, and what it is capable of doing, it’s worth considering a few different scenarios when you might want to use a Bokeh Server.

Local or Individual Use

One way that you might want to use the Bokeh server is during exploratory data analysis, possibly in a Jupyter notebook. Alternatively, you might want to create a small app that you can run locally, or that you can send to colleagues to run locally. The Bokeh server is very useful and easy to use in this scenario. Both of the methods here below can be used effectively:

For the most flexible approach, that could transition most directly to a deployable application, it is suggested to follow the techniques in Building Bokeh Applications.

Creating Deployable Applications

Another way that you might want to use the Bokeh server is to publish interactive data visualizations and applications that can be viewed and used by a wider audience (perhaps on the internet, or perhaps on an internal company network). The Bokeh Server is also well-suited to this usage, and you will want to first consult the sections:

Shared Publishing

Both of the scenarios above involve a single creator making applications on the server, either for their own local use, or for consumption by a larger audience. Another scenario is the case where a group of several creators all want publish different applications to the same server. This is not a good use-case for single Bokeh server. Because it is possible to create applications that execute arbitrary python code, process isolation and security concerns make this kind of shared tenancy prohibitive.

In order to support this kind of multi-creator, multi-application environment, one approach is to build up infrastructure that can run as many Bokeh servers as-needed, either on a per-app, or at least a per-user basis. It is possible that we may create a public service to enable just this kind of usage in the future, and it would also certainly be possible for third parties to build their own private infrastructure to do so as well, but that is beyond the scope of this User’s Guide.

Another possibility is to have a single centrally created app (perhaps by an organization), that can access data or other artifacts published by many different people (possibly with access controls). This sort of scenario is possible with the Bokeh server, but often involves integrating a Bokeh server with other web application frameworks. See a complete example at https://github.com/bokeh/bokeh-demos/tree/master/happiness

Building Bokeh Applications

By far the most flexible way to create interactive data visualizations using the Bokeh server is to create Bokeh Applications, and serve them with the bokeh serve command.

Single module format

Let’s look again at a complete example and then examine some specific parts in more detail:

# myapp.py

from random import random

from bokeh.layouts import column
from bokeh.models import Button
from bokeh.palettes import RdYlBu3
from bokeh.plotting import figure, curdoc

# create a plot and style its properties
p = figure(x_range=(0, 100), y_range=(0, 100), toolbar_location=None)
p.border_fill_color = 'black'
p.background_fill_color = 'black'
p.outline_line_color = None
p.grid.grid_line_color = None

# add a text renderer to out plot (no data yet)
r = p.text(x=[], y=[], text=[], text_color=[], text_font_size="20pt",
           text_baseline="middle", text_align="center")

i = 0

ds = r.data_source

# create a callback that will add a number in a random location
def callback():
    global i

    # BEST PRACTICE --- update .data in one step with a new dict
    new_data = dict()
    new_data['x'] = ds.data['x'] + [random()*70 + 15]
    new_data['y'] = ds.data['y'] + [random()*70 + 15]
    new_data['text_color'] = ds.data['text_color'] + [RdYlBu3[i%3]]
    new_data['text'] = ds.data['text'] + [str(i)]
    ds.data = new_data

    i = i + 1

# add a button widget and configure with the call back
button = Button(label="Press Me")
button.on_click(callback)

# put the button and plot in a layout and add to the document
curdoc().add_root(column(button, p))

Notice that we have not specified an output or connection method anywhere in this code. It is a simple script that creates and updates objects. The flexibility of the bokeh command line tool means that we can defer output options until the end. We could, e.g., run bokeh json myapp.py to get a JSON serialized version of the the application. But in this case, we would like to run the app on a Bokeh server, so we execute:

bokeh serve --show myapp.py

The --show option will cause a browser to open up a new tab automatically to the address of the running application, which in this case is:

http://localhost:5006/myapp

If you have only one application, the server root will redirect to it. Otherwise, You can see an index of all running applications at the server root:

http://localhost:5006/

This index can be disabled with the --disable-index option, and the redirect behavior can be disabled with the --disable-index-redirect option.

In addition to creating Bokeh applications from single python files, it is also possible to create applications from directories.

Directory format

Bokeh applications may also be created by creating and populating a filesystem directory with the appropriate files. To start a directory application in a directory myapp, execute bokeh serve with the name of the directory, for instance:

bokeh serve --show myapp

At a minimum, the directory must contain a main.py that constructs a Document for the Bokeh Server to serve:

myapp
   |
   +---main.py

The full set of files that Bokeh server knows about is:

myapp
   |
   +---main.py
   +---server_lifecycle.py
   +---static
   +---theme.yaml
   +---templates
        +---index.html

The optional components are

  • A server_lifecycle.py file that allows optional callbacks to be triggered at different stages of application creation, as descriped in Lifecycle Hooks.
  • A static subdirectory that can be used to serve static resources associated with this application.
  • A theme.yaml file that declaratively defines default attributes to be applied to Bokeh model types.
  • A templates subdirectory with index.html Jinja template file. The directory may contain additional Jinja templates for index.html to refer to. The template should have the same parameters as the FILE template.

When executing your main.py Bokeh server ensures that the standard __file__ module attribute works as you would expect. So it is possible to include data files or custom user defined models in your directory however you like. Additionally, the application directory is also added to sys.path so that python modules in the application directory may be easily imported.

An example might be:

myapp
   |
   +---data
   |    +---things.csv
   |
   +---helpers.py
   +---main.py
   |---models
   |    +---custom.js
   |
   +---server_lifecycle.py
   +---static
   |    +---css
   |    |    +---special.css
   |    |
   |    +---images
   |    |    +---foo.png
   |    |    +---bar.png
   |    |
   |    +---js
   |        +---special.js
   |
   |---templates
   |    +---index.html
   |
   +---theme.yaml

In this case you might have code similar to:

from os.path import dirname, join
from helpers import load_data

load_data(join(dirname(__file__), 'data', 'things.csv')

And similar code to load the JavaScript implementation for a custom model from models/custom.js

Accessing the HTTP Request

When a session is created for a Bokeh application, the session context is made available as curdoc().session_context. The most useful function of the session context is to make the Tornado HTTP request object available to the application as session_context.request. The request object has a number of fields, such as arguments, cookies, protocol, etc. See the documentation for HTTPServerRequest for full details.

As an example, the following code will access the request arguments to set a value for a variable N (perhaps controlling the number of points in a plot):

# request.arguments is a dict that maps argument names to lists of strings,
# e.g, the query string ?N=10 will result in {'N': [b'10']}

args = curdoc().session_context.request.arguments

try:
  N = int(args.get('N')[0])
except:
  N = 200

Warning

The request object is provided so that values such as arguments may be easily inspected. Calling any of the Tornado methods such as finish() or writing directly to request.connection is unsupported and will result in undefined behavior.

Callbacks and Events

Before jumping in to callbacks and events specifically in the context of the Bokeh Server, it’s worth discussing different use-cases for callbacks in general.

JavaScript Callbacks in the Browser

Regardless of whether there is a Bokeh Server involved, it is possible to create callbacks that execute in the browser, using CustomJS and other methods. See Div for more detailed information and examples.

It is critical to note that no python code is ever executed when a CustomJS callback is used. This is true even when the call back is supplied as python code to be translated to JavaScript. A CustomJS callback is only executed inside a browser JavaScript interpreter, and can only directly interact JavaScript data and functions (e.g., BokehJS Backbone models).

Python Callbacks with Jupyter Interactors

If you are working in the Jupyter Notebook, it is possible to use Jupyter interactors to quickly create simple GUI forms automatically. Updates to the widgets in the GUI can trigger python callback functions that execute in the Jupyter Python kernel. It is often useful to have these callbacks call push_notebook() to push updates to displayed plots. For more detailed information, see Jupyter Interactors.

Note

It is currently possible to push udpates from python, to BokehJS (i.e., to update plots, etc.) using push_notebook(). It is not currently possible to get events or updates from the other direction (e.g. to have a range or selection update trigger a python callback) without using a Bokeh Server as described in the next section. Adding the capability for two-way Python<–>JS synchronization through Jupyter comms is a planned future addition.

Updating From Threads

If the app needs to perform blocking computation, it can be possible to have a separate thread perform that work, and then add a callback to update the document with the results. It is important to emphasize that the interface to update the document must pass through a “next tick callback”. A callback added this way will execute as soon as possible on the next iteration of the Tornado event loop, and automatically acquire necessary locks to update the document state safely.

Any usage that updates the document state from another thread, either by calling other methods on the document, or by setting properties directly on Bokeh models, risks data and protocol corruption.

Warning

The ONLY safe operations to perform on a document from a different thread is add_next_tick_callback() and remove_next_tick_callback()

It is also important to save a local copy of curdoc() off so that all threads have access to the same document. This is illustrated in the example below:

from functools import partial
from random import random
from threading import Thread
import time

from bokeh.models import ColumnDataSource
from bokeh.plotting import curdoc, figure

from tornado import gen

# this must only be modified from a Bokeh session allback
source = ColumnDataSource(data=dict(x=[0], y=[0]))

# This is important! Save curdoc() to make sure all threads
# see then same document.
doc = curdoc()

@gen.coroutine
def update(x, y):
    source.stream(dict(x=[x], y=[y]))

def blocking_task():
    while True:
        # do some blocking computation
        time.sleep(0.1)
        x, y = random(), random()

        # but update the document from callback
        doc.add_next_tick_callback(partial(update, x=x, y=y))

p = figure(x_range=[0, 1], y_range=[0,1])
l = p.circle(x='x', y='y', source=source)

doc.add_root(p)

thread = Thread(target=blocking_task)
thread.start()

To see this example in action, save it to a python file, e.g. testapp.py and then execute

bokeh serve --show testapp.py

Warning

There is currently no locking around adding next tick callbacks to documents. It is recommended that at most one thread add callbacks to the document. It is planned to add more fine grained locking to callback methods in the future.

Updating from Unlocked Callbacks

You may also want to drive blocking computations from callbacks using, e.g. Tornado’s ThreadPoolExecutor in an asynchronous callback. This can work, however, normally Bokeh session callbacks recursively lock the document until all future work they initiate is completed. To make this scenario work as desired, Bokeh provides a without_document_lock() decorator that can suppress the normal locking behavior.

As with the thread example above, all actions that update document state must go through a next-tick callback.

The following example demonstrates an application that drives a blocking computation from one unlocked Bokeh session callback, by yielding to a blocking function that runs on the thread pool executor and updates by using a next-tick callback, and also updates the state simply from a standard locked session callback on a different update rate.

from functools import partial
import time

from concurrent.futures import ThreadPoolExecutor
from tornado import gen

from bokeh.document import without_document_lock
from bokeh.models import ColumnDataSource
from bokeh.plotting import curdoc, figure

source = ColumnDataSource(data=dict(x=[0], y=[0], color=["blue"]))

i = 0

doc = curdoc()

executor = ThreadPoolExecutor(max_workers=2)

def blocking_task(i):
    time.sleep(1)
    return i

# the unlocked callback uses this locked callback to safely update
@gen.coroutine
def locked_update(i):
    source.stream(dict(x=[source.data['x'][-1]+1], y=[i], color=["blue"]))

# this unclocked callback will not prevent other session callbacks from
# executing while it is in flight
@gen.coroutine
@without_document_lock
def unlocked_task():
    global i
    i += 1
    res = yield executor.submit(blocking_task, i)
    doc.add_next_tick_callback(partial(locked_update, i=res))

@gen.coroutine
def update():
    source.stream(dict(x=[source.data['x'][-1]+1], y=[i], color=["red"]))

p = figure(x_range=[0, 100], y_range=[0,20])
l = p.circle(x='x', y='y', color='color', source=source)

doc.add_periodic_callback(unlocked_task, 1000)
doc.add_periodic_callback(update, 200)
doc.add_root(p)

As before, you can run this example by saving to a python file and running bokeh serve on it.

Lifecycle Hooks

Sometimes it is desirable to have code execute at specific times in a server or session lifetime. For instance, if you are using a Bokeh Server along side a Django server, you would need to call django.setup() once, as each Bokeh server started, to initialize the Django properly for use by Bokeh application code.

Bokeh provides this capability through a set of Lifecycle Hooks. To use these hooks, you must create your application in Directory format, and include a designated file called server_lifecycle.py in the directory. In this file you can include any or all of the following conventionally named functions:

def on_server_loaded(server_context):
    ''' If present, this function is called when the server first starts. '''
    pass

def on_server_unloaded(server_context):
    ''' If present, this function is called when the server shuts down. '''
    pass

def on_session_created(session_context):
    ''' If present, this function is called when a session is created. '''
    pass

def on_session_destroyed(session_context):
    ''' If present, this function is called when a session is closed. '''
    pass

Connecting with bokeh.client

With the new Tornado and websocket-based server introduced in Bokeh 0.11, there is also a proper client API for interacting directly with a Bokeh Server. This client API can be used to trigger updates to the plots and widgets in the browser, either in response to UI events from the browser or as a results of periodic or asynchronous callbacks. As before, the first step is to start a Bokeh Server:

bokeh serve

Next, let’s look at a complete example, and then examine a few key lines individually:

import numpy as np
from numpy import pi

from bokeh.client import push_session
from bokeh.driving import cosine
from bokeh.plotting import figure, curdoc

x = np.linspace(0, 4*pi, 80)
y = np.sin(x)

p = figure()
r1 = p.line([0, 4*pi], [-1, 1], color="firebrick")
r2 = p.line(x, y, color="navy", line_width=4)

# open a session to keep our local document in sync with server
session = push_session(curdoc())

@cosine(w=0.03)
def update(step):
    # updating a single column of the the *same length* is OK
    r2.data_source.data["y"] = y * step
    r2.glyph.line_alpha = 1 - 0.8 * abs(step)

curdoc().add_periodic_callback(update, 50)

session.show(p) # open the document in a browser

session.loop_until_closed() # run forever

If you run this script, you will see a plot with an animated line appear in a new browser tab. The first half of the script is like most any script that uses the bokeh.plotting interface. The first interesting line is:

session = push_session(curdoc())

This line opens a new session with the Bokeh Server, initializing it with our current Document. This local Document will be automatically kept in sync with the server. The next few lines define and add a periodic callback to be run every 50 milliseconds:

@cosine(w=0.03)
def update(step):
    # updating a single column of the the *same length* is OK
    r2.data_source.data["y"] = y * step
    r2.glyph.line_alpha = 1 - 0.8 * abs(step)

curdoc().add_periodic_callback(update, 50)

Next, analogous to bokeh.io.show(), there is this a show() on session objects that will automatically open a browser tab to display the synced Document.

Finally, we need to tell the session to loop forever, so that the periodic callbacks happen:

session.loop_until_closed() # run forever

This mode of interaction can be very useful, especially for individual exploratory data analysis (e.g, in a Juypter notebook). However, it does have some drawbacks when compared to the Application technique described below. In particular, in addition to network traffic between the browser and the server, there is network traffic between the python client and the server as well. Depending on the particular usage, this could be a significant consideration.

Deployment Scenarios

With an application, we can run it just locally any time we want to interact with it. Or we can share it with other people, and they can run it locally themselves in the same manner. But we might also want to deploy the application in a way that other people can access it. This section describes some of the considerations that arise in that case.

Standalone Bokeh Server

First, it is possible to simply run the Bokeh server on a network for users to interact with directly. Depending on the computational burden of your application code, the number of users, the power of the machine used to run on, etc., this could be a simple and immediate option for deployment an internal network.

However, it is often the case that there are needs around authentication, scaling, and uptime. In these cases more sophisticated deployment configurations are needed. In the following sections we discuss some of these considerations.

SSH Tunnels

It may be convenient or necessary to run a standalone instance of the Bokeh server on a host to which direct access cannot be allowed. In such cases, ssh can be used to “tunnel” to the server.

In the simplest scenario, the Bokeh server will run on one host and will be accessed from another location, e.g., a laptop, with no intermediary machines.

Run the server as usual on the remote host:

bokeh server

Next, issue the following command on the local machine to establish an ssh tunnel to the remote host:

ssh -NfL localhost:5006:localhost:5006  user@remote.host

Replace user with your username on the remote host and remote.host with the hostname/IP address of the system hosting the Bokeh server. You may be prompted for login credentials for the remote system. After the connection is set up you will be able to navigate to localhost:5006 as though the Bokeh server were running on the local machine.

The second, slightly more complicated case occurs when there is a gateway between the server and the local machine. In that situation a reverse tunnel must be estabished from the server to the gateway. Additionally the tunnel from the local machine will also point to the gateway.

Issue the following commands on the remote host where the Bokeh server will run:

nohup bokeh server &
ssh -NfR 5006:localhost:5006 user@gateway.host

Replace user with your username on the gateway and gateway.host with the hostname/IP address of the gateway. You may be prompted for login credentials for the gateway.

Now set up the other half of the tunnel, from the local machine to the gateway. On the local machine:

ssh -NfL localhost:5006:localhost:5006 user@gateway.host

Again, replace user with your username on the gateway and gateway.host with the hostname/IP address of the gateway. You should now be able to access the Bokeh server from the local machine by navigating to localhost:5006 on the local machine, as if the Bokeh server were running on the local machine. You can even set up client connections from a Jupyter notebook running on the local machine.

Note

We intend to expand this section with more guidance for other tools and configurations. If have experience with other web deployment scenarios and wish to contribute your knowledge here, please contact us on the mailing list.

Basic Reverse Proxy Setup

If the goal is to serve an web application to the general Internet, it is often desirable to host the application on an internal network, and proxy connections to it through some dedicated HTTP server. This sections provides guidance for basic configuration behind some common reverse proxies.

Nginx

One very common HTTP and reverse-proxying server is Nginx. A sample server confuguration block is shown below:

server {
    listen 80 default_server;
    server_name _;

    access_log  /tmp/bokeh.access.log;
    error_log   /tmp/bokeh.error.log debug;

    location / {
        proxy_pass http://127.0.0.1:5100;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection "upgrade";
        proxy_http_version 1.1;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header Host $host:$server_port;
        proxy_buffering off;
    }

}

The above server block sets up Nginx to to proxy incoming connections to 127.0.0.1 on port 80 to 127.0.0.1:5100 internally. To work in this configuration, we will need to use some of the command line options to configure the Bokeh Server. In particular we need to use --port to specify that the Bokeh Server should listen itself on port 5100. We also need to set the --host option to whitelist 127.0.0.1:80 as an acceptable Host on the incoming request header:

bokeh serve myapp.py --port 5100 --host 127.0.0.1:80

Note

The --host option is to guard against spoofed Host values. In a more realistic scenario where you have Nginx and the Bokeh server server running on foo.com, you would set --host foo.com:80. Then any attempted connections that do not report this Host in the request header (as all connections from Nginx do) will be rejected.

Note that in the basic server block above we have not configured any special handling for static resources, e.g., the Bokeh JS and CSS files. This means that these files are served directly by the Bokeh server itself. While this works, it places an unnecessary additional load on the Bokeh server, since Nginx has a fast static asset handler. To utilize Nginx to server Bokeh’s static assets, you can add a new stanza inside the server block above, similar to this:

location /static {
    alias /path/to/bokeh/server/static;
}

Be careful that the file permissions of the Bokeh resources are accessible to whatever user Nginx is running as. Alternatively, you can copy the resources to a global static directory during your deployment process. See A Full Example with Automation for a demonstration of this.

Apache

Another common HTTP server and proxy is Apache:

<VirtualHost *:80>
    ServerName localhost

    CustomLog "/path/to/logs/access_log" combined
    ErrorLog "/path/to/logs/error_log"

    ProxyPreserveHost On
    ProxyPass /myapp/ws ws://127.0.0.1:5100/myapp/ws
    ProxyPassReverse /myapp/ws ws://127.0.0.1:5100/myapp/ws

    ProxyPass /myapp http://127.0.0.1:5100/myapp/
    ProxyPassReverse /myapp http://127.0.0.1:5100/myapp/

    <Directory />
        Require all granted
        Options -Indexes
    </Directory>

    Alias /static /path/to/bokeh/server/static
    <Directory /path/to/bokeh/server/static>
        # directives to effect the static directory
        Options +Indexes
    </Directory>

</VirtualHost>

The above configuration aliases /static to the location of the Bokeh static resources directory, however it is also possible (and probably preferable) to copy the Bokeh static resources to whatever standard static files location is configured for Apache as part of the deployment.

As before, you would run the Bokeh server with the command:

bokeh serve myapp.py --port 5100 --host 127.0.0.1:80

Reverse Proxying with Nginx and SSL

If you would like to deploy a Bokeh Server behind an SSL-terminated Nginx proxy, then a few additional customizations are needed. First, the Bokeh server must be configured for a --host with the HTTP port 443, and you must also add the --use-xheaders flag:

bokeh serve myapp.py --port 5100 --host foo.com:443 --use-xheaders

The --use-xheaders option causes Bokeh to override the remote IP and URI scheme/protocol for all requests with X-Real-Ip, X-Forwarded-For, X-Scheme, X-Forwarded-Proto headers when they are available.

You must also customize Nginx. In particular, you must configure Nginx to send the X-Forwarded-Proto header, as well as configure Nginx for SSL termination. Optionally, you may want to redirect all HTTP traffic to HTTPS. The complete details of this configuration (e.g. how and where to install SSL certificates and keys) will vary by platform, but a reference nginx.conf is provided below:

# redirect HTTP traffic to HTTPS (optional)
server {
    listen      80;
    server_name foo.com;
    return      301 https://$server_name$request_uri;
}

server {
    listen      443 default_server;
    server_name foo.com;

    # add Strict-Transport-Security to prevent man in the middle attacks
    add_header Strict-Transport-Security "max-age=31536000";

    ssl on;

    # SSL installation details will vary by platform
    ssl_certificate /etc/ssl/certs/my-ssl-bundle.crt;
    ssl_certificate_key /etc/ssl/private/my_ssl.key;

    # enables all versions of TLS, but not SSLv2 or v3 which are deprecated.
    ssl_protocols TLSv1 TLSv1.1 TLSv1.2;

    # disables all weak ciphers
    ssl_ciphers "ECDHE-RSA-AES256-GCM-SHA384:ECDHE-RSA-AES128-GCM-SHA256:DHE-RSA-AES256-GCM-SHA384:DHE-RSA-AES128-GCM-SHA256:ECDHE-RSA-AES256-SHA384:ECDHE-RSA-AES128-SHA256:ECDHE-RSA-AES256-SHA:ECDHE-RSA-AES128-SHA:DHE-RSA-AES256-SHA256:DHE-RSA-AES128-SHA256:DHE-RSA-AES256-SHA:DHE-RSA-AES128-SHA:ECDHE-RSA-DES-CBC3-SHA:EDH-RSA-DES-CBC3-SHA:AES256-GCM-SHA384:AES128-GCM-SHA256:AES256-SHA256:AES128-SHA256:AES256-SHA:AES128-SHA:DES-CBC3-SHA:HIGH:!aNULL:!eNULL:!EXPORT:!DES:!MD5:!PSK:!RC4";

    ssl_prefer_server_ciphers on;

    location / {
        proxy_pass http://127.0.0.1:5100;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection "upgrade";
        proxy_http_version 1.1;
        proxy_set_header X-Forwarded-Proto $scheme;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header Host $host:$server_port;
        proxy_buffering off;
    }

}

This configuration will proxy all incoming HTTPS connections to foo.com to a Bokeh server running internally on http://127.0.0.1:5100.

Load Balancing with Nginx

The architecture of the Bokeh server is specifically designed to be scalable—by and large, if you need more capacity, you simply run additional servers. Often in this situation it is desired to run all the Bokeh server instances behind a load balancer, so that new connections are distributed amongst the individual servers.

Nginx offers a load balancing capability. We will describe some of the basics of one possible configuration, but please also refer to the Nginx load balancer documentation. For instance, there are various different strategies available for choosing what server to connect to next.

First we need to add an upstream stanze to our NGinx configuration, typically above the server stanza. This section looks something like:

upstream myapp {
    least_conn;                 # Use Least Connections strategy
    server 127.0.0.1:5100;      # Bokeh Server 0
    server 127.0.0.1:5101;      # Bokeh Server 1
    server 127.0.0.1:5102;      # Bokeh Server 2
    server 127.0.0.1:5103;      # Bokeh Server 3
    server 127.0.0.1:5104;      # Bokeh Server 4
    server 127.0.0.1:5105;      # Bokeh Server 5
}

We have labeled this upstream stanza as myapp. We will use this name below. Additionally, we have listed the internal connection information for six different Bokeh server instances (each running on a different port) inside the stanza. You can run and list as many Bokeh servers as you need.

You would run the Bokeh servers with commands similar to:

serve myapp.py --port 5100 --host 127.0.0.1:80
serve myapp.py --port 5101 --host 127.0.0.1:80
...

Next, in the location stanza for our Bokeh server, change the proxy_pass value to refer to the upstream stanza we created above. In this case we use proxy_pass http://myapp; as shown here:

server {

    location / {
        proxy_pass http://myapp;

        # all other settings unchanged
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection "upgrade";
        proxy_http_version 1.1;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header Host $host:$server_port;
        proxy_buffering off;
    }

}

Process Control with Supervisord

It is often desired to use process control and monitoring tools when deploying web applications. One popular such tool is Supervisor, which can automatically start and stop process, as well as re-start processes if they terminate unexpectedly. Supervisor is configured using INI style config files. A sample file that might be used to start a single Bokeh Server app is below:

; supervisor config file

[unix_http_server]
file=/tmp/supervisor.sock   ; (the path to the socket file)
chmod=0700                  ; sockef file mode (default 0700)

[supervisord]
logfile=/var/log/supervisord.log ; (main log file; default $CWD/supervisord.log)
pidfile=/var/run/supervisord.pid ; (supervisord pidfile; default $CWD/supervisord.pid)
childlogdir=/var/log/supervisor  ; ('AUTO' child log dir, default $TEMP)

; The section below must be in the present for the RPC (supervisorctl/web)
; interface in to function.
[rpcinterface:supervisor]
supervisor.rpcinterface_factory = supervisor.rpcinterface:make_main_rpcinterface

[supervisorctl]
serverurl=unix:///tmp/supervisor.sock ; use a unix:// URL for a unix socket

[program:myapp]
command=/path/to/bokeh serve myapp.py --host foo.com:80
directory=/path/to/workdir
autostart=false
autorestart=true
startretries=3
numprocs=4
process_name=%(program_name)s_%(process_num)02d
stderr_logfile=/var/log/myapp.err.log
stdout_logfile=/var/log/myapp.out.log
user=someuser
environment=USER="someuser",HOME="/home/someuser"

The standard location for the supervisor configj file varies from system to system. Consult the Supervisor configuration documentation for more details. It is also possible to specify a config file explicity. To do this, execute:

supervisord -c /path/to/supervisord.conf

to start the Supervisor process. Then to control processes execute supervisorctl commands. For instance to start all processes, run:

supervisorctl -c /path/to/supervisord.conf start all

To stop all processes run:

supervisorctl -c /path/to/supervisord.conf start all

And to update the process control after editing the config file, run:

supervisorctl -c /path/to/supervisord.conf update

Scaling the server

You can fork multiple server processes with the num-procs option. For example, to fork 3 processes:

bokeh serve --num-procs 3

Note that the forking operation happens in the underlying Tornado Server, see notes in the Tornado docs.

A Full Example with Automation

To deploy the demo site at http://demo.bokeh.org we combine all of the above techniques. Additionally, we used SaltStack to automate many aspects of the deployment.

Note

Other devops automation tools include Puppet, Ansible, and Chef. We would like to provide specific guidance where ever we can, so if you have experience with these tools and would be interested in contributing your knowledge, please contact us on the mailing list.

You can see all the code for deploying the site at the public GitHub repository here:

https://github.com/bokeh/demo.bokeh.org

You can modify or deploy your own version of this site on an Amazon Linux instance by simply running the deploy.sh script at the top level. With minor modifications, this machinery should work on many linux variants.