Tune concurrency

Tune concurrency#

This guide shows you how to size the limiters of a ploneasgi deployment.

Set the worker count#

PLONE_CONCURRENCY bounds concurrent WSGI calls and native-route calls together, and each one holds a ZODB connection. Keep it at or below the ZODB pool-size in zope.conf, which defaults to 7:

PLONE_CONCURRENCY=6

To raise the pool, set pool-size in the database section of zope.conf, or db_pool_size when you generate the instance with cookiecutter-zope-instance.

check_concurrency_contract() logs a warning at startup when the worker count exceeds the pool.

Downloads do not need workers, so do not raise this number for file traffic. Raise it for slow pages: renders that wait on a remote service or a slow storage.

Set the body-read limiter#

Concurrent body reads have their own limiter, which defaults to four times the workers, and at least 16. To change it, pass your own adapter to create_app:

from ploneasgi.wsgi import WSGIMiddleware


def adapter(wsgi_app, settings):
    return WSGIMiddleware(
        wsgi_app,
        workers=settings.concurrency,
        io_workers=64,
        chunk_size=settings.chunk_size,
        max_body_size=settings.max_body_size or None,
    )


app = create_app(settings, wsgi_app=wsgi_app, adapter=adapter)

Keep workers=settings.concurrency: native routes share that limiter.

Set the chunk size#

PLONE_BLOB_CHUNK_SIZE is the number of bytes pulled from a file per thread hop, 256 KiB by default. Raise it for faster transfers of large files; each in-flight download holds one chunk in memory.

PLONE_BLOB_CHUNK_SIZE=1048576

Size the blob cache on a networked storage#

On ZEO, RelStorage, and PGJsonb, the first download of a blob fetches it into a local cache on a worker. Make the blob cache large enough to hold the files that are downloaded often, so that the fetch happens once. With cookiecutter-zope-instance, the option is db_blob_cache_size, or db_pgjsonb_blob_cache_size for PGJsonb.

Check the result#

Turn on the access log with PLONE_ACCESS_LOG=1 and watch ttfb under load. ttfb climbing while duration stays flat means requests are waiting for a worker.

Contents