Precompute-related updates

[tomato42]

• little bit of cleanup after the changes around precomputation code
• add CI on python 3.9 as it's released now
• more documentation
• more test coverage

[coveralls]

Coverage increased (+0.03%) to 98.028% when pulling 0dc17f0 on tomato42:precompute-updates into 73f3d3e on warner:master.

[tomato42]

CI results: https://travis-ci.org/github/warner/python-ecdsa/builds/742510210

[letoams]

Looks good to me. My only question is what the the precompute has a lazy flag at all. Why not just always have it lazy? Is there a use case where being lazy would be slower?

[tomato42]

yes, the first operation with lazy precompute will be much slower (like few hundred times slower), so if the code is latency sensitive, it may be too late to do it the first time the key is actually used

[letoams]

yes, the first operation with lazy precompute will be much slower (like few hundred times slower), so if the code is latency sensitive, it may be too late to do it the first time the key is actually used

But is that really a factor ? If performance were that critical, would you use python? Wouldn't this need to use ctypes ?

If you startup a program to do calculations, then you always take this hit anyway before you can "get to work". So the only use case we are talking about is a daemon waiting to do signing operations. Is "a few hundred times slower" meaningful compared to like 20ms of network latency ?

It feels a bit more like feature creep than actual real life use case to me. But having said that, I don't object to any of it.

[tomato42]

yes, the first operation with lazy precompute will be much slower (like few hundred times slower), so if the code is latency sensitive, it may be too late to do it the first time the key is actually used

But is that really a factor ? If performance were that critical, would you use python? Wouldn't this need to use ctypes ?

because on very underperforming platforms (like Raspberry Pi) that few hundred times slower translates for few seconds, see #211 and linked issues

If you startup a program to do calculations, then you always take this hit anyway before you can "get to work". So the only use case we are talking about is a daemon waiting to do signing operations. Is "a few hundred times slower" meaningful compared to like 20ms of network latency ?

yes, on a regular PC it's just 20ms so it doesn't matter, but on a phone it may be 2-3 seconds

It feels a bit more like feature creep than actual real life use case to me. But having said that, I don't object to any of it.

well, yeah, but it's reusing the code we need to library init so... meh