POTHEAD

POTHEAD

What?

POTHEAD uses a reverse-http proxy solution to improve request-latency when load-balancing expensive non-concurrent HTTP requests.

Why?

A certain class of http-backend-requests are poorly served by regular HTTP-load-balancing solutions, whether hashed or round-robin. This class of requests cannot efficiently over-use resources in the worker, for example due to breaking RAM-limits, or concurrency causing non-optimal CPU cache use. In a traditional forwarding HTTP load-balancer, the worker can throttle incoming requests by slowing down "accept"-rate, but doing so would increase latency and potentially leave free workers unused. One prime example is transcoders of audio, video or images, which is typically CPU-intensive and cache-sensitive.

How?

POTHEAD solves this problem by employing "reverse"-HTTP on the worker side. The TCP "client" (the worker initiating the TCP-connection), implements the server side of the HTTP, protocol, waiting for the TCP "server" to initiate the HTTP request. Both the workers and the service consumers connect to a service hub. Requests from the consumers are queued by the hub and dequed when a worker connects. The worker can thus control how many parallel connections to maintain, thereby the concurrency of the requests.

Why not?

To control the concurrency, the worker might need to employ Connection: close in order to accept a new request only when resources are available. This TCP reconnection leads to some overhead in network traffic, latency, and could lead to the TCP "lingering" problem. Therefore it's not recommended to use POTHEAD for requests with less than 50ms of average execution time.

Prometheus metrics

If the env variables PROMETHEUS_MULTIPROC_DIR is set to an existing directory it will be wiped and used for prometheus client in multiprocess mode. The variable PROMETHEUS_PORT can be used to change metrics export web server port (default 9090).

Usage in worker

Usage is the standard prometheus-client usage:

from prometheus_client import Counter

REQUEST_TOTAL = Counter(
    'requests_total',
    'Total HTTP requests',
    ['method', 'endpoint']
)

def app(environ, start_response):
    REQUEST_TOTAL.labels(
        method=environ['REQUEST_METHOD'],
        endpoint=environ['PATH_INFO']
    ).inc()
    start_response("200 OK", [('Content-Type','text/plain; charset=utf-8')])
    return ['hello'.encode('utf-8')];

(Why "POTHEAD"?)

Because PTTH was taken.

Ok, ok. How do I get started?

This implementation provides a hub based on aiohttp. It will open up two ports, one main port for consumers and one for workers. Run with python3 -m pothead.server.

It also includes a WSGI-enabled worker-runner, allowing you to host your regular WSGI-app through POTHEAD. Run using python3 -m pothead.worker --connect <host>:<port> <module>:<app-symbol>.

The runner have a couple of useful features, one being a gating-based "--poll-jobs" mode, where a wait_for_slot implemented on the provided app-object allows the application to dynamically pull jobs matching based on available resources. A standard implementation for CPU-usage-based gating is provided in pothead.gating.

Another worker-feature worth mentioning is "--redirect-response". Running in this mode, the worker will automatically and transparently redirect any successful (200) responses from this WSGI-app, to a direct port of the worker. This is useful to avoid the PTTH-broker becoming a bottleneck of network bandwidth.

Run tests with tox. If you're on MacOS and the build fails, use a Docker container: docker build -f dev.Dockerfile -t pothead-dev . && docker run -it -v $PWD:$PWD -w $PWD pothead-dev.

Err, could you show me some UML?

Redirection mode:

@startuml
skinparam maxMessageSize 250
participant Client as C
participant "PTTH Broker" as PB
participant "PTTH Worker" as PW
participant "WSGI App" as WW

PW -> PW : Start PTTH worker with WSGI app
PW -> PW : Create a <i>Server</i>, configuring <i>wait_for_slot</i> from the WSGI app.
PW -> PW : Wrap the WSGI app inside the <i>Server</i> in an <i>OutOfBandResponder</i> and start a server listening on the redirect port
PW -> WW : <i>Server</i> polls <i>wait_for_slot</i>
WW --> PW : Slot is available
activate PW #dddddd
activate PB #aaaaaa
PW -> PB : Connect over TCP
activate C
C -> PB : HTTP Request
PB -> PB : Pair Client Request with worker connection
PB -> PW : Proxy Client Request
activate WW
PW -> WW : <i>OutOfBandResponder</i>: Pass Client request to WSGI app
WW -> WW : Validate request
WW --> PW : 200 OK, chunked response
PW -> PW : Observe 200 OK, and capture a generator over the subsequent chunks of response into a map keyed by a generated <i>response key</i>
PW --> PB : <i>OutOfBandResponder</i>: HTTP 303 redirect to <i><PTTH worker ip>:<redirect_port>/<response key></i>
deactivate PW
PB --> C : HTTP 303 redirect
deactivate PB
activate PW #ffffff
C -> PW : Follow HTTP 303 redirect
PW -> PW : <i>OutOfBandResponder</i>: Look up response using <i><response key></i> from the request path
WW --> PW : Continue capturing chunks from WSGI app response
PW --> C : <i>OutOfBandResponder</i>: Respond 200 OK, and forward the captured chunks of response from WSGI app
@enduml

What then?

How would I know? You tell me.

License

Copyright 2019 Ulrik Mikaelsson

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.