Web3 Internals


This section of the documentation is for advanced users. You should probably stay away from these APIs if you don’t know what you are doing.

The Web3 library has multiple layers of abstraction between the public api exposed by the web3 object and the backend or node that web3 is connecting to.

  • Providers are responsible for the actual communication with the blockchain such as sending JSON-RPC requests over HTTP or an IPC socket.

  • Middlewares provide hooks for monitoring and modifying requests and responses to and from the provider. These can be global operating on all providers or specific to one provider.

  • Managers provide thread safety and primitives to allow for asynchronous usage of web3.

Here are some common things you might want to do with these APIs.

  • Redirect certain RPC requests to different providers such as sending all read operations to a provider backed by Infura and all write operations to a go-ethereum node that you control.

  • Transparently intercept transactions sent over eth_sendTransaction, sign them locally, and then send them through eth_sendRawTransaction.

  • Modify the response from an RPC request so that it is returned in different format such as converting all integer values to their hexadecimal representation.

  • Validate the inputs to RPC requests

Request Lifecycle

Each web3 RPC call passes through these layers in the following manner.

  ***********    ************
  | Request |    | Response |
  ***********    ************
      |                ^
      v                |
|            Manager          |
      |                ^
      v                |
|     Global Middlewares      |
      |                ^
      v                |
|    Provider Middlewares     |
      |                ^
      v                |
|          Provider           |

You can visualize this relationship like an onion, with the Provider at the center. The request originates from the Manager, outside of the onion, passing down through each layer of the onion until it reaches the Provider at the center. The Provider then handles the request, producing a response which will then pass back out from the center of the onion, through each layer until it is finally returned by the Manager.


A provider is responsible for all direct blockchain interactions. In most cases this means interacting with the JSON-RPC server for an ethereum node over HTTP or an IPC socket. There is however nothing which requires providers to be RPC based, allowing for providers designed for testing purposes which use an in-memory EVM to fulfill requests.

Writing your own Provider

Writing your own provider requires implementing two required methods as well as setting the middlewares the provider should use.

BaseProvider.make_request(method, params)

Each provider class must implement this method. This method should return a JSON object with either a 'result' key in the case of success, or an 'error' key in the case of failure.

  • method This will be a string representing the JSON-RPC method that is being called such as 'eth_sendTransaction'.

  • params This will be a list or other iterable of the parameters for the JSON-RPC method being called.


This function should return True or False depending on whether the provider should be considered connected. For example, an IPC socket based provider should return True if the socket is open and False if the socket is closed.

If set to True, the optional show_traceback boolean will raise a ProviderConnectionError and provide information on why the provider should not be considered connected.


This should be an iterable of middlewares.

You can set a new list of middlewares by assigning to provider.middlewares, with the first middleware that processes the request at the beginning of the list.



The Middleware API in web3 borrows heavily from the Django middleware API introduced in version 1.10.0

Middlewares provide a simple yet powerful api for implementing layers of business logic for web3 requests. Writing middleware is simple.

def simple_middleware(make_request, w3):
    # do one-time setup operations here

    def middleware(method, params):
        # do pre-processing here

        # perform the RPC request, getting the response
        response = make_request(method, params)

        # do post-processing here

        # finally return the response
        return response
    return middleware

It is also possible to implement middlewares as a class.

class SimpleMiddleware:
    def __init__(self, make_request, w3):
        self.w3 = w3
        self.make_request = make_request

    def __call__(self, method, params):
        # do pre-processing here

        # perform the RPC request, getting the response
        response = self.make_request(method, params)

        # do post-processing here

        # finally return the response
        return response

The make_request parameter is a callable which takes two positional arguments, method and params which correspond to the RPC method that is being called. There is no requirement that the make_request function be called. For example, if you were writing a middleware which cached responses for certain methods your middleware would likely not call the make_request method, but instead get the response from some local cache.

The RequestManager object exposes the middleware_onion object to manage middlewares. It is also exposed on the Web3 object for convenience. That API is detailed in Configuring Middleware.


The Manager acts as a gatekeeper for the request/response lifecycle. It is unlikely that you will need to change the Manager as most functionality can be implemented in the Middleware layer.

Request Processing for Persistent Connection Providers

class web3.providers.websocket.request_processor.RequestProcessor

The RequestProcessor class is responsible for the storing and syncing up of asynchronous requests to responses for a PersistentConnectionProvider. The best example of one such provider is the WebsocketProviderV2. In order to send a websocket message and receive a response to that particular request, PersistentConnectionProvider instances have to match request id values to response id values coming back from the websocket connection. Any provider that does not adhere to the JSON-RPC 2.0 specification in this way will not work with PersistentConnectionProvider instances. The specifics of how the request processor handles this are outlined below.

Listening for Responses

Implementations of the PersistentConnectionProvider class have a message listener background task that is called when the websocket connection is established. This task is responsible for listening for any and all messages coming in over the websocket connection and storing them in the RequestProcessor instance internal to the PersistentConnectionProvider instance. The RequestProcessor instance is responsible for storing the messages in the correct cache, either the one-to-one cache or the one-to-many (subscriptions) queue, depending on whether the message has a JSON-RPC id value or not.

One-To-One Requests

One-to-one requests can be summarized as any request that expects only one response back. An example is using the eth module API to request the latest block number.

>>> async def wsV2_one_to_one_example():
...     async with AsyncWeb3.persistent_websocket(
...         WebsocketProviderV2(f"ws://")
...     ) as w3:
...         # make a request and expect a single response returned on the same line
...         latest_block_num = await w3.eth.block_number

>>> asyncio.run(wsV2_one_to_one_example())

With websockets we have to call send() and asynchronously receive responses via another means, generally by calling recv() or by iterating on the websocket connection for messages. As outlined above, the PersistentConnectionProvider class has a message listener background task that handles the receiving of messages.

Due to this asynchronous nature of sending and receiving, in order to make one-to-one request-to-response calls work, we have to save the request information somewhere so that, when the response is received, we can match it to the original request that was made (i.e. the request with a matching id to the response that was received). The stored request information is then used to process the response when it is received, piping it through the response formatters and middlewares internal to the web3.py library.

In order to store the request information, the RequestProcessor class has an internal RequestInformation cache. The RequestInformation class saves important information about a request.

class web3._utils.caching.RequestInformation

The name of the method - e.g. “eth_subscribe”.


The params used when the call was made - e.g. (“newPendingTransactions”, True).


The formatters that will be used to process the response.


Any middleware that processes responses that is present on the instance at the time of the request is appended here, in order, so the response may be piped through that logic when it comes in.


If the request is an eth_subscribe request, rather than popping this information from the cache when the response to the subscription call comes in (i.e. the subscription id), we save the subscription id with the request information so that we can correctly process all subscription messages that come in with that subscription id. For one-to-one request-to-response calls, this value is always None.

One-to-one responses, those that include a JSON-RPC id in the response object, are stored in an internal SimpleCache class, isolated from any one-to-many responses. When the PersistentConnectionProvider is looking for a response internally, it will expect the message listener task to store the response in this cache. Since the request id is used in the cache key generation, it will then look for a cache key that matches the response id with that of the request id. If the cache key is found, the response is processed and returned to the user. If the cache key is not found, the operation will time out and raise a TimeExhausted exception. This timeout can be configured by the user when instantiating the PersistentConnectionProvider instance via the response_timeout keyword argument.

One-To-Many Requests

One-to-many requests can be summarized by any request that expects many responses as a result of the initial request. The only current example is the eth_subscribe request. The initial eth_subscribe request expects only one response, the subscription id value, but it also expects to receive many eth_subscription messages if and when the request is successful. For this reason, the original request is considered a one-to-one request so that a subscription id can be returned to the user on the same line, but the process_subscriptions() method on the WebsocketConnection class, the public API for interacting with the active websocket connection, is set up to receive eth_subscription responses over an asynchronous interator pattern.

>>> async def ws_v2_subscription_example():
...     async with AsyncWeb3.persistent_websocket(
...         WebsocketProviderV2(f"ws://")
...     ) as w3:
...         # Subscribe to new block headers and receive the subscription_id.
...         # A one-to-one call with a trigger for many responses
...         subscription_id = await w3.eth.subscribe("newHeads")
...         # Listen to the websocket for the many responses utilizing the ``w3.ws``
...         # ``WebsocketConnection`` public API method ``process_subscriptions()``
...         async for response in w3.ws.process_subscriptions():
...             # Receive only one-to-many responses here so that we don't
...             # accidentally return the response for a one-to-one request in this
...             # block
...             print(f"{response}\n")
...             if some_condition:
...                 # unsubscribe from new block headers, another one-to-one request
...                 is_unsubscribed = await w3.eth.unsubscribe(subscription_id)
...                 if is_unsubscribed:
...                     break

>>> asyncio.run(ws_v2_subscription_example())

One-to-many responses, those that do not include a JSON-RPC id in the response object, are stored in an internal asyncio.Queue instance, isolated from any one-to-one responses. When the PersistentConnectionProvider is looking for one-to-many responses internally, it will expect the message listener task to store these messages in this queue. Since the order of the messages is important, the queue is a FIFO queue. The process_subscriptions() method on the WebsocketConnection class is set up to pop messages from this queue as FIFO over an asynchronous iterator pattern.

If the stream of messages from the websocket is not being interrupted by any other tasks, the queue will generally be in sync with the messages coming in over the websocket. That is, the message listener will put a message in the queue and the process_subscriptions() method will pop that message from the queue and yield control of the loop back to the listener. This will continue until the websocket connection is closed or the user unsubscribes from the subscription. If the stream of messages lags a bit, or the provider is not consuming messages but has subscribed to a subscription, this internal queue may fill up with messages until it reaches its max size and then trigger a waiting asyncio.Event until the provider begins consuming messages from the queue again. For this reason, it’s important to begin consuming messages from the queue, via the process_subscriptions() method, as soon as a subscription is made.