git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 23/23] initial pack index v3 support on the read side

From
Nicolas Pitre <nico@fluxnic.net>
Date
Sep 3, 2013, 06:09 UTC
Message-ID
<alpine.LFD.2.03.1309030201230.14472@syhkavp.arg>
In-Reply-To
<CACsJy8CzvQSYRkH+FNuNhzicX9CLC7bRxKgq2RDmOuxh1Q-=5Q@mail.gmail.com>
On Sat, 31 Aug 2013, Duy Nguyen wrote:
Show 17 quoted lines
> On Tue, Aug 27, 2013 at 11:26 AM, Nicolas Pitre <nico@fluxnic.net> wrote:
> > A bit crud but good enough for now.
> 
> I wonder if we should keep a short SHA-1 table in .idx. An entry in
> the original .idx v1 table is <SHA-1>+<offset> then offset moved out
> to make the table more compact for binary search. I observe that we
> don't always need 20 byte SHA-1 to uniquely identify an entry in a
> pack, so the SHA-1 table could be split in two: one table contain the
> first part of SHA-1, long enough to identify any entry in the pack;
> the second table contains either full SHA-1 or the remaining part.
> Binary search is done on the first table, if matched, full sha-1 from
> the second table is compared. We already have the second sha-1 table
> in .pack v4. We could add the first table in .idx v3.
> 
> On linux-2.6 even in one-pack configuration, we only need the first 6
> bytes of sha-1 to identify an object. That would cut the bsearch sha-1
> table to 1/4 full sha-1 table size.
I don't see the point though.
Why having two tables when only one suffice?

If the argument is about making the SHA1 binary search more efficient given a smaller memory footprint, that comes with extra complexity coming from the variable length SHA1 strings in the second table. So I'm not sure there is much to gain.

Furthermore, one of the design of pack v4 is to avoid the SHA1 binary search entirely. You will need the binary search only once to locate the top commit object, but from there the entire object tree can be walked simply by using the sha1ref index and looking up the corresponding offset into the pack index file to locate the next object.

Same thing for walking tree objects: the pack v4 tree deltas are meant to be walked inline without having to _apply_ any delta.

Nicolas
Previous: Duy NguyenNext: Duy Nguyen
Message 48 of 83 in “Preliminary pack v4 support”
  1. 00/23 Preliminary pack v4 supportNicolas Pitre, Aug 27, 2013
  2. 01/23 pack v4: initial pack dictionary structure and codeNicolas Pitre, Aug 27, 2013
  3. Junio C HamanoAug 27, 2013
  4. Nicolas PitreAug 27, 2013
  5. 02/23 export packed_object_info()Nicolas Pitre, Aug 27, 2013
  6. 03/23 pack v4: scan tree objectsNicolas Pitre, Aug 27, 2013
  7. 04/23 pack v4: add tree entry mode support to dictionary entriesNicolas Pitre, Aug 27, 2013
  8. 05/23 pack v4: add commit object parsingNicolas Pitre, Aug 27, 2013
  9. Junio C HamanoAug 27, 2013
  10. Nicolas PitreAug 27, 2013
  11. Junio C HamanoAug 27, 2013
  12. 06/23 pack v4: split the object list and dictionary creationNicolas Pitre, Aug 27, 2013
  13. 07/23 pack v4: move to struct pack_idx_entry and get rid of our own struct idx_entryNicolas Pitre, Aug 27, 2013
  14. 08/23 pack v4: basic references encodingNicolas Pitre, Aug 27, 2013
  15. Junio C HamanoAug 27, 2013
  16. Nicolas PitreAug 27, 2013
  17. 09/23 pack v4: commit object encodingNicolas Pitre, Aug 27, 2013
  18. Junio C HamanoAug 27, 2013
  19. Nicolas PitreAug 27, 2013
  20. Nicolas PitreAug 27, 2013
  21. Junio C HamanoAug 27, 2013
  22. Nicolas PitreAug 27, 2013
  23. Duy NguyenSep 2, 2013
  24. Nicolas PitreSep 3, 2013
  25. Duy NguyenSep 3, 2013
  26. Nicolas PitreSep 5, 2013
  27. 10/23 pack v4: tree object encodingNicolas Pitre, Aug 27, 2013
  28. Junio C HamanoAug 27, 2013
  29. Nicolas PitreAug 27, 2013
  30. 11/23 pack v4: dictionary table outputNicolas Pitre, Aug 27, 2013
  31. 12/23 pack v4: creation codeNicolas Pitre, Aug 27, 2013
  32. Junio C HamanoAug 27, 2013
  33. Nicolas PitreAug 27, 2013
  34. 13/23 pack v4: object headersNicolas Pitre, Aug 27, 2013
  35. 14/23 pack v4: object data copyNicolas Pitre, Aug 27, 2013
  36. Junio C HamanoAug 27, 2013
  37. Nicolas PitreAug 27, 2013
  38. 15/23 pack v4: object writingNicolas Pitre, Aug 27, 2013
  39. 16/23 pack v4: tree object delta encodingNicolas Pitre, Aug 27, 2013
  40. 17/23 pack v4: load delta candidate for encoding tree objectsNicolas Pitre, Aug 27, 2013
  41. 18/23 pack v4: honor pack.compression config optionNicolas Pitre, Aug 27, 2013
  42. 19/23 pack v4: relax commit parsing a bitNicolas Pitre, Aug 27, 2013
  43. 20/23 pack index v3Nicolas Pitre, Aug 27, 2013
  44. 21/23 pack v4: normalize pack name to properly generate the pack index file nameNicolas Pitre, Aug 27, 2013
  45. 22/23 pack v4: add progress displayNicolas Pitre, Aug 27, 2013
  46. 23/23 initial pack index v3 support on the read sideNicolas Pitre, Aug 27, 2013
  47. Duy NguyenAug 31, 2013
  48. Nicolas PitreSep 3, 2013
  49. Duy NguyenSep 3, 2013
  50. Document pack v4 formatNguyễn Thái Ngọc Duy, Aug 27, 2013
  51. Junio C HamanoAug 27, 2013
  52. Nicolas PitreAug 27, 2013
  53. Document pack v4 formatNguyễn Thái Ngọc Duy, Aug 31, 2013
  54. Nicolas PitreSep 3, 2013
  55. Nicolas PitreSep 3, 2013
  56. Duy NguyenSep 3, 2013
  57. Duy NguyenSep 3, 2013
  58. Nicolas PitreSep 5, 2013
  59. Duy NguyenSep 5, 2013
  60. Nicolas PitreSep 5, 2013
  61. Duy NguyenSep 5, 2013
  62. Nicolas PitreSep 5, 2013
  63. Duy NguyenSep 5, 2013
  64. Nicolas PitreSep 5, 2013
  65. Junio C HamanoSep 5, 2013
  66. Nicolas PitreSep 5, 2013
  67. Duy NguyenSep 6, 2013
  68. Nicolas PitreSep 6, 2013
  69. Document pack v4 formatNguyễn Thái Ngọc Duy, Sep 6, 2013
  70. Nicolas PitreSep 6, 2013
  71. Duy NguyenSep 6, 2013
  72. Nicolas PitreSep 6, 2013
  73. Duy NguyenSep 6, 2013
  74. Nicolas PitreSep 6, 2013
  75. Nicolas PitreSep 7, 2013
  76. Nicolas PitreSep 7, 2013
  77. Duy NguyenSep 7, 2013
  78. Junio C HamanoAug 27, 2013
  79. Nicolas PitreAug 27, 2013
  80. Junio C HamanoAug 27, 2013
  81. Duy NguyenAug 28, 2013
  82. Nicolas PitreAug 28, 2013
  83. Duy NguyenAug 28, 2013

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.