Create 2019-04-07--ipns.md

meeting-notes/2019/Q2/2019-04-17--ipns.md

@@ -0,0 +1,106 @@
+# Semi-Centralized IPNS Status April 17, 2019
+
+- **Lead:** @aschmahmann
+- **Notetaker:**
+  - @momack2
+- **Attendees:**
+  - @aschmahmann
+  - @satazor
+  - @vasco-santos
+  - @hugomrdias
+  - @jimpick
+  - @lidel
+  - @momack2
+- [**Zoom.us meeting URL**](https://protocol.zoom.us/j/868858751)
+- [**Recording**](https://drive.google.com/open?id=1SFrWN_PJsTZYe_VkKhUHxCMa-4fzL2tW)
+
+## Agenda
+
+1. Start recording
+2. Ask everyone to put their name into the list of attendees
+3. Ask for a volunteer to take notes
+4. Everyone can add items to this agenda for things they would like to discuss
+
+5. Collect list of current proposals/ongoing work (IPFS DNSlink, Cloudflare workers, IPNS-over-DNS, IPNS-over-Pubsub,...)
+6. What happens when we have multiple IPNS networks? (i.e. we need to specify ipns:// the protocol)
+7. Re-publish what can we do ?
+
+## Notes
+
+- Hugo
+  - Investigating short-term solution for IPFS camp where a fast IPNS would support fast IPFS-on-wb-browers and also package manager work
+  - doing proof of concepts using cloudflare workers
+    - it's distributed because it runs on the edges, but not fully decentralized
+    - did it in 1-2 days, easy to do
+    - 150ms round trip time to IPNS
+    - works like an API, the workers get the requests and either put record in key-value store or you get record from there
+    - when put validate record is valid, and then self-certify when get it
+    - not much trust of the code running in the workers
+    - biggest problem is if workers are blacklisted by cloudflare / removed, hard to replicate the database in other servers (need to change things like http addresses)
+    - think can get down to 100ms globally
+  - other option is to pick up IPNS over DNS
+    - https://github.com/ipfs/RFC/compare/rfc/5-ipns-over-dns
+    - think more likely to be accepted by community in the short term
+    - RFC made by lars, but when update DNSlink record on your domain, put IPNS record with domain instead of with hash
+    - make the "resolve" part mostly done - resolves recursively until finds the IPFS
+    - already working in go, still a little more js work
+    - publishing a little harder, need an authoritive server for some domain that is attributed to IPNS - then people do IPNS path through that domain until it resolves to an IPFS path
+    - could tell people they could deploy on their own servers, config name servers to point to their own servers
+    - with DNS already have global infra to rely on so when it works, it's fast
+    - "how does DNS figure our which servers are the authoritative ones?"
+      - IPFS provides one resolve (ex dweb.link), and if people use hash.dweb.link, know it points to desired thing
+      - only have to do this once even when have multiple versions - point the hash.dweb.link 
+    - resolver will always listen on a well-known pubsub channel and when publish can go to that pubsub channel to publish updats
+      - resolver saves that
+      - won't necessarily need a domain for things like package managers because can use ipns path directly (websites need to have domain)
+      - important to be able to update directly via IPNS instead of constantly updating the DNS record (risky and doesn't allow things like github previews)
+    - DNS is fast to update, but the problem is TTLs, caches, and routers
+    - when request bubbles up through the DNS record, will know the authoritive domain is the one in the record
+    - republishing is the biggest problem
+    - could default to "dweb.link" when not explicitly specified
+- Adin
+  - IPNS over pubsub - the current issue is it doesn't resolve on the initial hit - have to hit the DHT for initial resolve
+    - now want to ask peers directly and they echo it out every now and then
+    - working reasonably well, still needs some IPFS plumbing
+    - relative to DNS, going to be a little slower because have to do a DHT query to find the people interested in your address
+    - faster because don't need quorum, just need a person on the pubsub channel
+      - can be mitigated by things like rendezvous
+      - instead stores name map to peer id, push to you all the people you're looking for
+      - need to hit rendezvous and then the pubsub channel
+      - rendezvous not quite done yet, but the non-federated version seems to be closer
+    - things run in a couple of seconds, but the DHT adds a level of randomness for benchmarking (rendezvous)
+    - think can get around adversarial nodes if enough of them
+  - seems like might create more overhead with many many pubsub channels for every website you visit
+  - some folks use pre-configured pubsub as a performance improvement over a centralized service (could do the subscription after the first central lookup)
+- right now have IPNS over DHT and IPNS over pubsub
+  - pubsub currently both published to DHT and pulls from there initially
+  - when add more ways to recieve a record, changing what it means to know where to query for default behavior
+  - want to be able to expriment, and standardizing this a single prototype will constrain that
+  - the DNS service option centralized overhead on our gateways (that are already in pain)
+  - being central authority for IPNS take downs a downside - if public gateway blocked then this is broken for the centralized support
+  - it dramatically simplifies the infra needed to NOT hit our gateway (easy to point record at own server - not in the DHT)
+- what's the plan for something like ipld:// - how would you resolve that?
+  - thinking about it as an internal pathing scheme, not how you grab/store the blocks
+  - for browser integration, the browser would need to have that configured
+  - would prefer to leverage something that was already there
+- want to specify the ipns protocol as we add more things
+  - like to always have DHT fallback (as last resort)
+  - maybe apart from DHT you can pick whatever makes sense from use case?
+  - continue discussion in issues
+  - the order should be configurable - defining the order won't always fit everyone's use cases (in the short term, DNS is always faster)
+  - what do you do with failures/inconsistency with updates
+- Republishing
+  - unreasonable for people to rerun ipns publish every day for website publishing
+  - would a long TTL work? or 3rd party republishable
+  - should be able to keep a website alive even if can't publish a new version. should be able to keep the last version alive forever if people care - if you disappear your link to your website shouldn't go stale
+  - keeping an old signed record alive should work
+    - people publish content with low TTL, and other people who want it to be maintained will just republish to new record with longer ttl - doesn't prevent and just makes more annoying for users who want this
+    - in gossipsub can keep republishing, but in DHT you can't
+    - problem with setting TTL to infinity: many records in DHT, if trying to prove out freshness don't want the old record also floating around
+  - would it make sense for different implementations to have different end of lifes?
+    - currently the DHT won't actually support infinite TTL (drops it after 24hrs and doesn't allow republishing)
+      - there's a desire for freshness, but maybe we should allow republishing if within the TTL
+    - for DNS, infinite TTL would be less of an issue
+    - maybe the DHT would just drop them?
+AIs:
+- Adin: Will start issue in specs repo discussing this - if needed will set up another meeting