Some queries a client may want to execute against connected relays are prohibitively expensive, for example, in order to retrieve follower counts for a given pubkey, a client must query all kind-3 events referring to a given pubkey only to count them. The result may be cached, either by a client or by a separate indexing server as an alternative, but both options erode the decentralization of the network by creating a second-layer protocol on top of Nostr.
This NIP defines the verb `COUNT`, which accepts a subscription id and filters as specified in [NIP 01](01.md) for the verb `REQ`. Multiple filters are OR'd together and aggregated into a single count result.
In case a relay uses probabilistic counts, it MAY indicate it in the response with `approximate` key i.e. `{"count": <integer>, "approximate": <true|false>}`.
This is so it enables merging results from multiple relays and yielding a reasonable estimate of reaction counts, comment counts and follower counts, while saving many millions of bytes of bandwidth for everybody.
1. Upon receiving a filter, if it has a single `#e`, `#p`, `#a` or `#q` item, read its 32th ascii character as a nibble (a half-byte, a number between 0 and 16) and add `8` to it to obtain an `offset` -- in the unlikely case that the filter doesn't meet these conditions, set `offset` to the number `16`;
2. Initialize 256 registers to `0` for the HLL value;
That is all that has to be done on the relay side, and therefore the only part needed for interoperability.
On the client side, these HLL values received from different relays can be merged (by simply going through all the registers in HLL values from each relay and picking the highest value for each register, regardless of the relay).
And finally the absolute count can be estimated by running some methods I don't dare to describe here in English, it's better to check some implementation source code (also, there can be different ways of performing the estimation, with different quirks applied on top of the raw registers).
One could mine a pubkey with a certain number of zero bits in the exact place where the HLL algorithm described above would look for them in order to artificially make its reaction or follow "count more" than others. For this to work a different pubkey would have to be created for each different target (event id, followed profile etc). This approach is not very different than creating tons of new pubkeys and using them all to send likes or follow someone in order to inflate their number of followers. The solution is the same in both cases: clients should not fetch these reaction counts from open relays that accept everything, they should base their counts on relays that perform some form of filtering that makes it more likely that only real humans are able to publish there and not bots or artificially-generated pubkeys.
The value `hll` value must be the concatenation of the 256 registers, each being a uint8 value (i.e. a byte). Therefore `hll` will be a 512-character hex string.
