[PERF] [JMAP] Improve attachement download

[chibenwa]

Why?

Today attachment are handled as mime parts, so if I download 5 attachments on the same mail I would need to read its full content 5 times, parse it then decode the attachment.

This is innefficient because we have a projection for attachments : it could be a simple download of that projection...

This mismatch leads to:

• Needless reads on the objectStore (billed by cloud provider)
• Unecessary parsing on James side, very likely a good proportion of the JMAP workload

What?

Allign the blob id on the JMAP layer with the attachment id on the mailbox layer

That way it's a direct download.

How?

Long term known issue.

I tend to look at it by the "JMAP layer needs to use mailbox attachmentId" but it was very complex and involved a lot of coupling.

After discussing with @HoussemNasri I believe that the other way around might be easier: "mailbox layer can be using JMAP blobId as an attachement Id"

To do this we could:

• Put MessageStorer behind an interface (and rename it! AttachmentParser? )
• Provide an EmailBodyPart backed version of it.
• Add the attachmentId back into the ParsedAttachement and make message parser responsible of the allocation
• Upon part downloads lookup the attachment first, then fallback to blob parts.
• And realign the tests

Should be doable...

[chibenwa]

I started POCing around this.

A bit hacky but I got linagora/tmail-backend#1466 and apache#2602 that kind a demonstrate the idea.

Retrocompatible (forward and reverse) with current code!

[chibenwa]

linagora/james-gatling#183 I can be helped by developing the benchmarks for this.