6 ms·

This anti-pattern -- floats as the sole numeric type -- goes all the way back to BASIC in 1964. It's just as bad an idea now as it was then.

SamReidHughes 10y ago
The only concrete problem anybody's mentioned with floats is that they only have 53 bits of precision, and some people need their integers to go up to 64 bits, or more.
- sedachv 10y ago
  What are you going to do for currency amounts if you do not have ints ? Strings with numbers in them?
  - SamReidHughes 10y ago
    Floating point numbers handle integer values in the range -2^53..2^53 just fine.
    - jnordwick 10y ago
      You can definitely get around the currency issue by scaling yourself, but things like the timestamp issues (fractional seconds since the epoch with millisecond precision) are a little more problematic. You basically have to roll your own datetime format and lose any db support. Databases are about data after all. Why not have a rich way to describe it?
      - SamReidHughes 10y ago
        I think so too. But I don't think, say, not having a 32-bit integer type, and 16-bit and 8-bit integers, is a big problem, if you've got doubles. Maybe it's a nice-to-have. A 64 bit type or bigint would add real value.
        jnordwick 10y ago
        I think as far as number types, you NEED (in the you should be considered seriously broken even if you can hobble along without) to have: - a 64-bit float - a 64-bit int everything else can be emulated in code from there and you can play all the encoding games you want to save space when you don't have to use it. This covers 99.999% of the use cases you reasonably see. From there, I would argue, you should also have: - a bit and byte - 32-bit ints - 32-bit float - 128-bit int (once you add UUIDs you might as well make them full numerics citizens). If I were doing a modern database, I'd have the kitchen sink as a type full numerics from 8 to 128 bit both signed and unsigned int and all supported hardware float sizes. I'd probably even have a 512-bit AVX type just to see what people would use it for.
        Dylan16807 10y ago
        Why do you NEED a 64 bit int? Why not 32 bit? You're not storing pointers in a database. (And then you can implement your 32 bit int in terms of the 53 bits of int-capacity in a double.)
        jnordwick 10y ago
        datetimes, aggregations, etc. pretty much anything non-toy in the modern world requires 64 bit int. like I said. read NEED as in you can certainly hobble around without it (pick an appropriate offset and read your 53 bits in relation to it), but like a fracture in a leg, it's still considered broken.
        Dylan16807 10y ago
        What exactly do you mean by aggregations here? If you want a good timestamp you probably want a timespec, and that won't fit in 64 bits anyway. I don't see why it's easy to make 128+ bit types be composite, but you can't have 64 bit types be composite. Your database driver can take care of how the bits are packed on the wire. It really shouldn't be a big concern.
- overgard 10y ago
  Here's two actual concrete problems: efficiency and type safety. Indexing an array with "1.5" doesn't seem awesome, nor does using 8 bytes for a double where a 1 byte int would do fine.
  - SamReidHughes 10y ago
    A double with integral value in 127..127 already takes 2 bytes to store in rethinkdb (the first being a tag distinguishing between array, object, string, etc), compared to some random double's 9 bytes. The type safety advantages of distinguishing doubles from integers are pretty minimal in a database, because you'd need a schema for that anyway, and the benefits of that far outweighs the add-on benefits of having type checked query logic.

SamReidHughes 10y ago

The only concrete problem anybody's mentioned with floats is that they only have 53 bits of precision, and some people need their integers to go up to 64 bits, or more.

sedachv 10y ago
What are you going to do for currency amounts if you do not have ints ? Strings with numbers in them?
- SamReidHughes 10y ago
  Floating point numbers handle integer values in the range -2^53..2^53 just fine.
  - jnordwick 10y ago
    You can definitely get around the currency issue by scaling yourself, but things like the timestamp issues (fractional seconds since the epoch with millisecond precision) are a little more problematic. You basically have to roll your own datetime format and lose any db support. Databases are about data after all. Why not have a rich way to describe it?
    - SamReidHughes 10y ago
      I think so too. But I don't think, say, not having a 32-bit integer type, and 16-bit and 8-bit integers, is a big problem, if you've got doubles. Maybe it's a nice-to-have. A 64 bit type or bigint would add real value.
      - jnordwick 10y ago
        I think as far as number types, you NEED (in the you should be considered seriously broken even if you can hobble along without) to have: - a 64-bit float - a 64-bit int everything else can be emulated in code from there and you can play all the encoding games you want to save space when you don't have to use it. This covers 99.999% of the use cases you reasonably see. From there, I would argue, you should also have: - a bit and byte - 32-bit ints - 32-bit float - 128-bit int (once you add UUIDs you might as well make them full numerics citizens). If I were doing a modern database, I'd have the kitchen sink as a type full numerics from 8 to 128 bit both signed and unsigned int and all supported hardware float sizes. I'd probably even have a 512-bit AVX type just to see what people would use it for.
        Dylan16807 10y ago
        Why do you NEED a 64 bit int? Why not 32 bit? You're not storing pointers in a database. (And then you can implement your 32 bit int in terms of the 53 bits of int-capacity in a double.)
        jnordwick 10y ago
        datetimes, aggregations, etc. pretty much anything non-toy in the modern world requires 64 bit int. like I said. read NEED as in you can certainly hobble around without it (pick an appropriate offset and read your 53 bits in relation to it), but like a fracture in a leg, it's still considered broken.
        Dylan16807 10y ago
        What exactly do you mean by aggregations here? If you want a good timestamp you probably want a timespec, and that won't fit in 64 bits anyway. I don't see why it's easy to make 128+ bit types be composite, but you can't have 64 bit types be composite. Your database driver can take care of how the bits are packed on the wire. It really shouldn't be a big concern.
overgard 10y ago
Here's two actual concrete problems: efficiency and type safety. Indexing an array with "1.5" doesn't seem awesome, nor does using 8 bytes for a double where a 1 byte int would do fine.
- SamReidHughes 10y ago
  A double with integral value in 127..127 already takes 2 bytes to store in rethinkdb (the first being a tag distinguishing between array, object, string, etc), compared to some random double's 9 bytes. The type safety advantages of distinguishing doubles from integers are pretty minimal in a database, because you'd need a schema for that anyway, and the benefits of that far outweighs the add-on benefits of having type checked query logic.