add option to run DHT without providers

[whyrusleeping]

Some systems (like filecoin) don't need providers, and having the providers subsystem running is a moderate annoyance (its currently causing an absurd amount of bandwidth usage :/). Can we get an option to run the dht without providers?

[vyzo]

But filecoin does need providers for autorelay.

[whyrusleeping]

@vyzo well thats a great point...

Advice? It hurts to run a node right now because of dht traffic

[raulk]

IMO the solution is to allow for client-mode switches for each functionality:

• Peer routing: client mode allows you to find peers, but it doesn't keep a routing table.
• Provider system: client mode allows you to resolve providers, but you ignore requests to store provider records.
• Content routing: client mode allows you to lookup content, but you don't store any.

[raulk]

^^ in that manner, Filecoin nodes would run the provider subsystem in client-mode, peer routing in full mode, and no content routing.

[whyrusleeping]

@raulk the problem with that though is that there would be no nodes left to store the provider records for autorelay rendezvous :/

[anacrolix]

I'd like to know how this issue might affect choices in #265. If anyone would like to comment there on how they might envisage how this issue might affect the projected API around modes and such, or whether these will be disjoint parts of the API, thanks.

[Stebalien]

@whyrusleeping are you worried about sending or storing provider records? You can pass a custom router wrapper to bitswap that turns Provide calls into no-ops (I've done this while testing go-ipfs performance in the past).

[raulk]

@anacrolix done:
#265 (comment)

@Stebalien apparently Filecoin was seeing an issue with nodes receiving a lot of provider record traffic, even though they don’t use this feature. However, they can’t disable it altogether because autorelay uses it. There’s something at play here. I started investigating but the proofs parameter generation took ages and I context switched. On my plate for tomorrow.

[tobowers]

Was there any outcome from this PR... I think we're seeing something similar - our CPU getting eaten up by provide.

[whyrusleeping]

@tobowers i don't think its made progress, as a workaround, try passing a completely unrelated datastore to the dht (don't use the same datastore that backs your blockstore)