On Thu, Jun 27, 2019 at 8:42 PM Derrick Stolee <stolee@xxxxxxxxx> wrote: > > On 6/27/2019 9:24 AM, Jeff Hostetler wrote: > > On 6/27/2019 6:48 AM, Duy Nguyen wrote: > >> On Tue, Jun 25, 2019 at 7:40 PM Derrick Stolee <stolee@xxxxxxxxx> wrote: > >>> > >>> On 6/25/2019 6:29 AM, Duy Nguyen wrote: > >>>> On Tue, Jun 25, 2019 at 3:06 AM Jeff Hostetler <git@xxxxxxxxxxxxxxxxx> wrote: > >>>>> I'm curious how big these EWAHs will be in practice and > >>>>> how useful an array of integers will be (especially as the > >>>>> pretty format will be one integer per line). Perhaps it > >>>>> would helpful to have an extended example in one of the > >>>>> tests. > >>>> > >>>> It's one integer per updated entry. So if you have a giant index and > >>>> updated every single one of them, the EWAH bitmap contains that many > >>>> integers. > >>>> > >>>> If it was easy to just merge these bitmaps back to the entry (e.g. in > >>>> this example, add "replaced": true to entry zero) I would have done > >>>> it. But we dump as we stream and it's already too late to do it. > >>>> > >>>>> Would it be better to have the caller of ewah_each_bit() > >>>>> build a hex or bit string in a strbuf and then write it > >>>>> as a single string? > >>>> > >>>> I don't think the current EWAH representation is easy to read in the > >>>> first place. You'll probably have to run through some script to update > >>>> the main entries part and will have a much better view, but that's > >>>> pretty quick. If it's for scripts, then it's probably best to keep as > >>>> an array of integers, not a string. Less post processing. > >>> > >>> I don't think the intent is to dump the EWAH directly, but instead to > >>> dump a string of the uncompressed bitmap. Something like: > >>> > >>> "delete_bitmap" : "01101101101" > >>> > >>> instead of > >>> > >>> "delete_bitmap" : [ 0, 1, 1, 0, 1, 1, 0, 1, 1, 1, 0, 1 ] > >> > >> I get this part. But the numbers in the array were the position of the > >> set bits. It's not showing just the actual bit map. > >> > >> The same bitmap would be currently displayed as > >> > >> "delete_bitmap": [ 1, 2, 4, 5, 7, 8, 9, 11 ] > >> > >> And that maps back to the entry[1], entry[2], entry[4]... in the index > >> being deleted from the base index. So displaying as a real bit map > >> actually adds more work for both the reader and the tool because you > >> have to calculate the position either way. And it gets harder if the > >> bit you're intereted in is on the far right. > > > > > > Thanks for the clarification. That helps. > > Same here! We expect these to be much smaller than the full set, correct? For split-index, the number of 1 bits should be about the size of your working set, not the index size. In the normal case, then yes it should be much smaller. After a big merge or branch switch, it could get as big as the index. But I would hope the logic to re-split the index kicks in, which essentially empties these bitmaps. EWAH bitmap is also used in UNTR extension if I remember correctly. Those bitmaps may have as many bits as the directories you have in the index. > Thanks, > -Stolee > -- Duy
