problems on import(roaring) about featurebase HOT 5 CLOSED

i'm wondering if the roaring implementation needs to replace the 16bit containers with 32 bit containers. The idea behind those containers is to have half the bits for low and half for high. So the container should be 32 bits instead of 64 bits.

Thought?

from featurebase.

I pushed up a PR to actually return the error message: 🤦
#37

Can you rerun the import and send the error?
As far as the containers, if they have 32-bits then the array size or else they become really inefficient. Instead of 4,096 elements it'd probably need to go up to 67,108,864 (4K x 16K). If we left it at 4K then we'd need to allocate a huge array for the bitmap container to hold all the elements.
Do you have a failing test case for the issue?
Ben

On Friday, January 8, 2016 10:43 AM, tgruben <[email protected]> wrote:

i'm wondering if the roaring implementation needs to replace the 16bit containers with 32 bit containers. The idea behind those containers is to have half the bits for low and half for high. So the container should be 32 bits instead of 64 bits.Thought?—
Reply to this email directly or view it on GitHub.

from featurebase.

You right on the inefficiency. Looking at the roaring code it seems to chop up the 32 bit number into 16bit parts..a high and a low. Perhaps we split up into 4 16 bit chunks. HighHigh, HighLow, LowHigh, LowLow and do the search. Never needing the 64 bit ops..ie search64.

And I would search in this order(for it would be just the same performance as roaring for values< 2^32
LowHigh,LowLow,HighLow,HighHigh

How does that grab you?

Here is the error...
2016/01/09 08:18:06 import error: db=3, frame=b.n, slice=%!s(uint64=59), bits=5434, err=open storage: unmarshal storage: file=/Users/tgruben/.pilosa/3/b.n/59, err=checksum mismatch: exp=5c922157, got=9dc388c3
2

The error is in importing the attached file above in slice 59, i'm trying to get it narrowed down better than that, but I haven't found the culprit. Below is the profile ids of the failed "import"

pids.txt

from featurebase.

It seems like it would add a lot of complexity by adding 4 levels instead of just 2. Unless we had a lot of containers it doesn't seem like it would help much.

from featurebase.

@tgruben I found and fixed the issue: #38

from featurebase.