Caching requests

Browser caching

The Cache-Control: header supersedes previous caching headers (e.g. Expires). Modern browsers support Cache-Control. This is all we need.

Here is an example of how to use Cache-Control::

    pattern: /$YAMLURL/path       # Pick any pattern
    handler: FileHandler          # and handler
        path: $YAMLPATH/path              # Pass it any arguments
        headers:                          # Define HTTP headers
            Cache-Control: max-age=3600   # Keep page in browser cache for 1 hour (3600 seconds)

The cache is used by browsers as well as proxies. You can also specify these additional options:

Here are some typical Cache-Control headers. The durations given here are indicative. Change them based on your needs.

To reload ignoring the cache, press Ctrl-F5 on the browser. Below is a useful reference for cache-control checks (Google Dev Docs):

HTTP Cache Control

Server caching

The url: handlers accept a cache: key that defines caching behaviour. For example, this configuration at random generates random letters every time it is called:

    pattern: /$YAMLURL/random
    handler: FunctionHandler
        function: random.choice(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'])

But adding the cache: to this URL caches it the first time it is called. When random-cached is reloaded, the same letter is shown every time.

    pattern: /$YAMLURL/random-cached
    handler: FunctionHandler
        function: random.choice(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'])
    cache: true

Cache keys

The response from any handler is cached against a cache key. By default, this is the URL. But you can change this using the cache.key argument.

For example, cache-full-url?x=1 and cache-full-url?x=2 return different values because they cache the full URL. But cache-only-path?x=1 and cache-only-path?x=2 return the same value because they only cache the path.

    pattern: /$YAMLURL/cache-full-url
    handler: FunctionHandler
        function: random.choice(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'])
        key: request.uri          # This is the default cache key

    pattern: /$YAMLURL/cache-only-path
    handler: FunctionHandler
        function: random.choice(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'])
        key: request.path     # only use the request path, not arguments

The key can accept multiple values. The values can either be:

For example, this configuration caches based on the request URI and user. Each URI is cached independently for each user ID.

        key:                # Cache based on
            - request.uri     # the URL requested
            -         # and handler.current_user['id'] if it exists

Google, Facebook, Twitter and LDAP provide the attribute. DB Auth provides it if your user table has an id column. But you can use any other attribute instead of id – e.g. for Google, user.screen_name for Twitter, etc.

Cache expiry

You can specify a expiry duration. For example cache-expiry caches the response for 5 seconds.

    pattern: /$YAMLURL/cache-expiry
    handler: FunctionHandler
        function: random.choice(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j'])
            duration: 5             # Cache the request for 5 seconds

By default, the cache expires either after 10 years, or when the cache store runs out of space.

Cache status

By default, only requests that return a HTTP 200 or HTTP 304 status code are cached. You can cache other status codes via the status: configuration.

    pattern: /$YAMLURL/cache-errors
        status: [200, 404, 500]         # Cache all of these HTTP responses

Cache stores

Gramex provides an in-memory cache, but you can define your own cache in the root cache: section as follows:

    small-in-memory-cache:  # Define a name for the cache
        type: memory        # This is an in-memory cache
        size: 100000        # Just allow 100K of data in the cache

    big-disk-cache:         # Define a name for the cache
        type: disk          # This is an on-disk cache
        path: $YAMLPATH/.cache  # Location of the disk cache directory
        size: 1000000000    # Allow ~1GB of data in the cache

    distributed-cache:      # Define a name for the cache
        type: redis         # This is a redis cache
        path: localhost:6379:0 # Connection string for Redis instance
        size: 1000000000   # Allow ~1GB of data in the cache

By default, Gramex provides a cache called memory that has a 500 MB in-memory cache based on cachetools. When the size limit is reached, the least recently used items are discarded. This cache is used by To change its size, use:

    memory:                 # This is the name of the pre-defined Gramex cache
        type: memory        # This is an in-memory cache
        size: 5000000       # Just allow 5MB of data in the cache instead of 500 MB (default)

Disk caches are based on the diskcache library. When the size limit is reached, the oldest items are discarded. But disk caches are MUCH slower than memory caches, and defeat the purpose of data caching. Use this if your app is computation or query intensive, and you need to share the cache across different instances on the same server.

Redis cache allows multiple gramex instances to cache objects in a Redis server. This allows the same cache to be used across different servers.

Redis cache requires Redis 5.0 or later to be running. When the size limit is reached, the oldest items are discarded. (Note: the size limit is set for the Redis instance, not for a specific DB. So avoid using the same Redis instance for other apps.)

To use a different cache by default, specify a default: true against the cache. The last cache with default: true is used as the default cache.

        default: false      # Don't use memory as the default cache
        type: memory
        size: 1000000000    # Allow ~1GB of data in the cache
        default: true

Note: Persistent caches like disk and redis pickle Python objects. Some objects (e.g. Tornado templates) are not pickle-able. These caches do not cache such objects, ignoring them with a log error. Use memory cache if you need to cache pickle-able objects.

Your functions can access these caches from cache object in gramex.service. For example, the default in-memory Gramex cache is at gramex.service.cache['memory']. The disk cache above is at gramex.service.cache['big-disk-cache'].

The cache stores can be treated like a dictionary. They also support a .set() method which accepts an expire= parameter. For example:

from gramex import service      # Available only if Gramex is running
cache = service.cache['big-disk-cache']
cache['key'] = 'value'
cache['key']      # returns 'value'
del cache['key']  # clears the key
cache.set('key', 'value', expire=30)    # key expires in 30 seconds

Mixing Python versions

The cache implementation in Python 2 is different from Python 3 because:

This means that you cannot have Gramex instances on Python 2 and Python 3 share the same cache. (Gramex instances running the same Python version can share the cache.)

Cache static files

You can cache static files with both server and client side caching. For example, to cache the bower_components and assets directories, use this configuration:

    pattern: /$YAMLURL/(bower_components/.*|assets/.*)    # Map all static files
    handler: FileHandler
    path: $YAMLPATH/                              # from under this directory
        Cache-Control: "public, max-age=315360000"  # Cache for 10 years on the browser
    cache: true                                     # Also cache on the server

To force a refresh, append ?v=xx where xx is a new number. (The use of ?v= is arbitrary. You can use any query parameter instead of v.)

Data caching opens files and caches them unless they are changed. You can use this to load any type of file. For example:

import gramex.cache
data ='data.csv', encoding='utf-8')

This loads data.csv using pd.read_csv('data.csv', encoding='utf-8'). The next time this is called, if data.csv in unchanged, the cached results are returned.

You can also specify that the file is a CSV file by explicitly passing a 2nd parameter as 'csv'. For example:

data ='data.csv', 'csv', encoding='utf-8')

(v1.23 made the 2nd parameter optional. It was mandatory before then.)

The 2nd parameter can take the following values:

The 2nd parameter can also be a function like function(path, **kwargs). For example:

# Return file size if it has changed
file_size ='data.csv', lambda path: os.stat(path).st_size)

# Read Excel file. Keyword arguments are passed to pd.read_excel
data ='data.xlsx', pd.read_excel, sheet_name='Sheet1')

To transform the data after loading, you can use a transform= function. For example:

# After loading, return len(data)
row_count ='data.csv', 'csv', transform=len)

# Return multiple calculations
def transform(data):
    return {'count': len(data), 'groups': data.groupby('city')}
result ='data.csv', 'csv', transform=transform)

You can also pass a rel=True parameter if you want to specify the filename relative to the current folder. For example, if D:/app/ has this code:

conf ='config.yaml', 'yaml', rel=True)

… the config.yaml will be loaded from the same directory as the calling file, D:/app/, that is from D:/app/config.yaml.

To simplify creating callback functions, use gramex.cache.opener. This converts functions that accept a handle or string into functions that accept a filename. gramex.cache.opener opens the file and returns the handle to the function.

For example, to read using pickle.load, use:

loader = gramex.cache.opener(pickle.load)
data ='template.pickle', loader)

If your function accepts a string instead of a handle, add the read=True parameter. This passes the results of reading the handle instead of the handle. For example, to compute the MD5 hash of a file, use:

m = hashlib.md5
loader = gramex.cache.opener(m.update, read=True)
data ='template.txt', mode='rb', encoding=None, errors=None)

Query caching

gramex.cache.query returns SQL queries as DataFrames and caches the results. The next time it is called, the query re-runs only if required.

For example, take this slow query:

query = '''
    SELECT,, SUM(sales.value)
    FROM product, sales
    WHERE = sales.product_id
    GROUP BY (,

If sales data is updated daily, we need not run this query unless the latest date has changed. Then we can use:

data = gramex.cache.query(query, engine, state='SELECT MAX(date) FROM sales')

gramex.cache.query is just like pd.read_sql but with an additional state= parameter. state can be a query – typically a fast running query. If running the state query returns a different result, the original query is re-run.

state can also be a function. For example, if a local file called .updated is changed every time the data is loaded, you can use:

data = gramex.cache.query(query, engine, state=lambda: os.stat('.updated').st_mtime)

Module caching

The Python import statement loads a module only once. If it has been loaded, it does not reload it.

During development, this means that you need to restart Gramex every time you change a Python file.

You can reload the module using six.moves.reload_module(module_name), but this reloads them module every time, even if nothing has changed. If the module has any large calculations, this slows things down.

Instead, use gramex.cache.reload_module(module_name). This is like six.moves.reload_module, but it reloads only if the file has changed.

For example, you can use it in a FunctionHandler:

import my_utils
import gramex.cache

def my_function_handler(handler):
    # Code used during development -- reload module if source has changed

You can use it inside a template:

{% import my_utils %}
{% import gramex.cache %}
{% set gramex.cache.reload_module(my_utils) %}
(Now my_utils.method() will have the latest saved code)

In both these cases, whenever is updated, the latest version will be used to render the FunctionHandler or template.

Subprocess streaming

You can run an OS command asynchronously using gramex.cache.Subprocess. Use this instead of subprocess.Popen because the latter will block Gramex until the command runs.

Basic usage:

def function_handler(handler):
    proc = gramex.cache.Subprocess(['python', '-V'])
    out, err = yield proc.wait_for_exit()
    # out contains stdout result. err contains stderr result
    # proc.proc.returncode contains the return code
    raise tornado.gen.Return('Python version is ' + err.decode('utf-8'))

out and err contain the stdout and stderr from running python -V as bytes. All keyword arguments supported by subprocess.Popen are supported here.

Streaming is supported. This lets you read the contents of stdout and stderr while the program runs. Example:

def function_handler(handler):
    proc = gramex.cache.Subprocess(['flake8'],
    out, err = yield proc.wait_for_exit()
    # out will be an empty byte string since stream_stdout is specified

This reads the output of flake8 line by line (since buffer_size='line') and writes the output by calling handler.write. The returned value for out is an empty string.

stream_stdout is a list of functions. You can provide any other method here. For example:

out = []
proc = gramex.cache.Subprocess(['flake8'],

… will write the output line-by-line into the out list using out.append.

stream_stderr works the same was as stream_stdout but on stderr instead.