Anne (:annevk) Reporter
	Description • 8 years ago

Currently we couple various domains together such as en.wikipedia.org and jp.wikipedia.org because they share a common public suffix (wikipedia.org) and give them a combined storage quota.

We do this to protect the user from someone creating many subdomains and filling the storage quota in each.

The alternative is creating many domains, which requires more cash and is therefore a less likely attack.

However, grouping subdomains together also makes it harder for larger sites to make use of storage, unless they register themselves as public suffix.

Since we have a maximum for storage and we use LRU to clear entries, should we really concern ourselves with attacks? Sure, if the user gets their storage attack once their next visit to favorite sites might go a little slower, but such attacks will also get flagged and banned (by malware protection and such) and have only a little bit of impact.

On the flipside, requiring legitimate use cases to go through the public suffix registry is onerous and some cannot even do that as it would mean their cookies would no longer be shared either, so they'd lose centralized login (this is the case for Wikipedia I believe).

I therefore think we should get rid of this limitation and just have a global quota and a per-origin quota.

	Anne (:annevk) Reporter
	Comment 1 • 7 years ago

Here's a concrete idea to improve LRU to make attacks harder: We introduce a two-step process when freeing up storage. First, we remove data older than X (e.g., six months or a year). Second, we remove data that was used the least amount of time. We calculate amount of time by subtracting first time something was stored from the last time storage was touched.

This way recent data that was used for a long time will likely not be removed and old data that was used for a long time back then will still go away.

Jan might be able to come up with a slightly more ambitious algorithm where we instead calculate a weight for each origin based on recency and time used.