From: Alex Ainscow Date: Thu, 5 Feb 2026 13:08:10 +0000 (+0000) Subject: docs: Update design document X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=5f29311db3f2d3cb5b8eb10bf89647e1e3aefdce;p=ceph.git docs: Update design document Signed-off-by: Alex Ainscow Signed-off-by: Callum James --- diff --git a/doc/dev/osd_internals/erasure_coding/direct_reads.rst b/doc/dev/osd_internals/erasure_coding/direct_reads.rst index 23337b24fa4..c0283b0190f 100644 --- a/doc/dev/osd_internals/erasure_coding/direct_reads.rst +++ b/doc/dev/osd_internals/erasure_coding/direct_reads.rst @@ -260,14 +260,16 @@ switched off: the bypass code must be trivially simple. Throttling Considerations ~~~~~~~~~~~~~~~~~~~~~~~~~ -This change presents a challenge for the existing throttling framework. It is -critical that the throttling behavior for a single logical client read remains -consistent, regardless of how many physical OSD operations it is split into. -The precise method for ensuring that the throttling scheduler correctly -accounts for these fragmented operations as a single logical unit is an open -problem that requires further investigation. As such, the design for this -component is not yet finalised and will be addressed during the implementation -phase. +The only place this code interacts with any throttling code is the Objecter +``throttle`` mechanism. This is a simple mechanism which collects two metrics +which it calls ``bytes``, derived from bytes transferred, and ``ops``. The ``SplitOp`` +mechanism will be inserted after the throttle processing has completed. No +attempt will be made to mark a direct reads as multiple ops. We can justify this +because: + +- Avoids changing client behavior too much. +- When ops are being split into direct reads, the bytes metric should dominate. +- This is the simpler approach. Replica Balanced Reads ----------------------