mesos-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Neil Conway <neil.con...@gmail.com>
Subject Re: Review Request 43635: Changed scalar resources to use fixed-point internally.
Date Wed, 17 Feb 2016 00:46:23 GMT

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/43635/#review119393
-----------------------------------------------------------




docs/attributes-resources.md (line 7)
<https://reviews.apache.org/r/43635/#comment180733>

    Not really, just a minor doc cleanup I made along the way. Happy to split into a separate
patch if you'd prefer.



docs/attributes-resources.md (line 39)
<https://reviews.apache.org/r/43635/#comment180734>

    Two spaces after a period is considered good style in English prose according to some
people (this style is used in various places throughout the comments and docs, but we aren't
consistent). I don't have a strong view, but it will require changing a lot more places than
just here to adopt a single style.



src/common/values.cpp (line 52)
<https://reviews.apache.org/r/43635/#comment180732>

    I'd prefer to keep the newline: without a newline, it suggests that the comment is specific
to the function that follows (`operator<<`), which would be misleading: the comment
applies to the following ~7 functions.



src/common/values.cpp (line 60)
<https://reviews.apache.org/r/43635/#comment180737>

    Truncation rather than rounding might also be reasonable behavior; I'm curious what other
people think.



src/common/values.cpp (line 61)
<https://reviews.apache.org/r/43635/#comment180736>

    I didn't use `setprecision`, because:
    
    1. There is value in making sure we use the same rounding method in this operator as elsewhere,
e.g, for corner-cases like 1.2345
    2. `setprecision` is a side-effecting operation, so modifying the caller's `ostream` seems
problematic.



src/common/values.cpp (line 67)
<https://reviews.apache.org/r/43635/#comment180735>

    What should we do in case of overflow?
    
    Note that we don't check for overflow in `operator+` (or for negative result values in
`operator-`)...


- Neil Conway


On Feb. 16, 2016, 11:29 p.m., Neil Conway wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/43635/
> -----------------------------------------------------------
> 
> (Updated Feb. 16, 2016, 11:29 p.m.)
> 
> 
> Review request for mesos, Joris Van Remoortere and Michael Park.
> 
> 
> Bugs: MESOS-4687
>     https://issues.apache.org/jira/browse/MESOS-4687
> 
> 
> Repository: mesos
> 
> 
> Description
> -------
> 
> Scalar resource values are represented using floating point. As a result, users
> could see unexpected results when accepting offers and making reservations for
> fractional resources: values like "0.1" cannot be precisely represented using
> standard floating point, and the resource values returned to frameworks might
> contain an unpredictable amount of roundoff error.
> 
> This commit adjusts the master to use fixed-point when doing internal
> computations on scalar resource values. The fixed-point format only supports
> three decimal digits of precision: that is, fractional resource values like
> "0.001" will be supported, but "0.0001" will not be.
> 
> 
> Diffs
> -----
> 
>   docs/attributes-resources.md 818da8ab0c672144b02f526b2b805cf0505d2c7e 
>   docs/upgrades.md 4f30d725c6ed28c09a1c5528fd4193c3f06b2d93 
>   include/mesos/mesos.proto e24d3e03a7dc7c6bfd07f34531cb593fe4925646 
>   include/mesos/v1/mesos.proto d909e60ddfd8e3ba2075f82c372edde04cd99d54 
>   src/common/resources.cpp 5d731870542166cec11f9956ccdf16207b2d22cc 
>   src/common/values.cpp c64407bc97ad858300f4661d616e0480920fc541 
>   src/master/allocator/mesos/hierarchical.cpp a9d2c23162892e22220f97d89a076d2311091d91

>   src/tests/resources_tests.cpp 96864c3945729fe33be8b243b9c826fb12e90eff 
>   src/v1/resources.cpp 207eb61d6a6d03d314539d42751cac65fcffa9af 
>   src/v1/values.cpp 58ea9875804bf0287855a1e9855855e5e54de4c4 
> 
> Diff: https://reviews.apache.org/r/43635/diff/
> 
> 
> Testing
> -------
> 
> make check
> 
> Manually verified that some of the floating point oddities in https://issues.apache.org/jira/browse/MESOS-4071
do not occur when this patch is applied, although I wasn't able to reproduce the crash described
in that ticket.
> 
> REVIEW NOTES:
> 
> * We don't currently emit a warning when discarding additional digits of precision from
input scalar resource values. Should we? That would require identifying all the points where
a resource value is seemed to be "user-provided", and also runs the risk of generating a ton
of log messages when an old framework is used.
> * Similarly, if the user gives us a resource value and we don't do anything to it, we
won't discard any additional precision that appears in the value -- the precision only gets
discarded when we apply an operator like `+` or `-`. Unclear if we should trim additional
precision from all scalar resource values more aggressively.
> 
> 
> Thanks,
> 
> Neil Conway
> 
>


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message