[OSM-talk] Facebook acquires crowdsourced mapping company Mapillary

Mark Wagner mark+osm at carnildo.com
Sun Jun 28 20:44:50 UTC 2020

On Sat, 27 Jun 2020 22:01:57 +0200
"Marc M." <marc_marc_irc at hotmail.com> wrote:

> Le 26.06.20 à 14:09, Florian Lohoff a écrit :
> >> I don't care about SLA. Does OSM have SLA?  
> > The point is when you distribute your storage to people at
> > home we will have at most 10% of images online all the time.  
> what facts are you basing that number on ?
> The worst internet connection I have has 98% availability.
> of course, it's not mandatory to have a pay-per-use dialup :)
> not to mention local chapters or companies with servers in datacenters
> or with fiber connection (where 2 instances of the PoC are running
> right now).
> > Disregarding the case that upstream bandwidth internationally
> > is pretty bad so you tend to have access times
> > for images of about 4-5 seconds at best. (3MByte image at
> > a typical ADSL upstream with 1.5MBit/s and international latencys)  
> your logic does not correspond to a distributed storage.
> it is not "one disk that sends one photo to one user"
> it is the pool that sends the pool of requests to all users.
> if you have 1000 adsl to serve 100 simultaneous requests,
> this is the equivalent of 15Mb/s par request (minus the management).
> I leave it up to you to imagine an order of magnitude for the
> conversion between simultaneous requests and users.

The only distributed data store I'm aware of with an actual
implementation is Freenet.  Freenet exhibits classic long-tail
behavior: popular data has high (>99%) availability with rapid
transfers, while the vast majority of data objects exist only as
references from other objects, and on the rare instances where they're
still in the data store, transfer times are measured in minutes.

If you treat a Bittorent index as a distributed data store, you see the
same pattern: a few popular torrents have hundreds or thousands of seeds
and download as fast as your connection can handle, while the majority
have, at best, one seeder, and many only have one or two users with
partial copies, hanging around in hopes that someone will eventually
show up with the remaining pieces.

In the context of a street-level photography system, distributed
storage means that you'll have no trouble downloading images of Old
Faithful or the Eiffel Tower, but you'll be lucky if you can get even
one photo of Road 783 in central Nebraska.

Distributed storage is one of those things that sounds good, but
nobody's figured out how to make it actually work well.


More information about the talk mailing list