1020554 - Mapper should yield content when returning full mapfiles

Status: RESOLVED WONTFIX

(Release Engineering :: General, defect)

Description

[Pete Moore [:pmoore][:pete]]

See https://bugzilla.mozilla.org/show_bug.cgi?id=847640#c30 for context.

Basically, mapper aggregates the full mapfile in memory, retrieving content from the database, and only when it has created the full response content in-memory, it starts to stream it back to the requester. Instead, it should write back to the http response as it gets data from the database, so there is not a significant delay between making a request, and starting to receive content.

This is important, since after 30s of inactivity, the load balancer will cut the connection and return an HTTP 500 response.

Comment 1

[Dustin J. Mitchell [:dustin] (he/him)]

We should verify, but I think that the load balancers give up after 30 seconds of inactivity, so as long as data is flowing, things are fine.

Flask does make streaming pretty easy, thankfully:

http://flask.pocoo.org/docs/patterns/streaming/

Comment 2

[hwine]

However, the mapfile data is sorted, so streaming may not reduce time-to-first-byte.

Comment 3

[Pete Moore [:pmoore][:pete]]

(In reply to Hal Wine [:hwine] (use needinfo) from comment #2)
> However, the mapfile data is sorted, so streaming may not reduce
> time-to-first-byte.

There is a delay from querying the DB until starting to receive rows back via the database connection, and there is a second delay from receiving the first data back from the database connection until the full mapfile is built up in memory. I think we can impact this second delay with streaming, but you are quite right Hal that the first part also incurs a delay. I'm not sure how much is the first part, and how much is the second part, at the moment.

Comment 4

[hwine]

no one to work on this at the moment

If anyone is seeing this in the real world, could they add details to help with prioritization.

Comment 5

[Pete Moore [:pmoore][:pete]]

No need to touch this. No problems in over a year of use, and not a system we need to heavily invest in, especially with deprioritisation of B2G which was main user.