Uploaded image for project: 'WiredTiger'
  1. WiredTiger
  2. WT-5477

Figure out problem in cursor caching for History Store cursors

    • Type: Icon: Improvement Improvement
    • Resolution: Duplicate
    • Priority: Icon: Major - P3 Major - P3
    • None
    • Affects Version/s: None
    • Component/s: None
    • None
    • 8
    • Storage Engines 2020-02-10

      While I work on WT-5467 to remove the cursor pool, I am having trouble at connection close with history store cursors. For some reason, history store cursors do not like to be cached as of the attached patch. I get the following crash if I do not disable cursor caching for the history store cursors:

      #0  __strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:93
      #1  0x00007fe0e0ea84d3 in _IO_vfprintf_internal (s=s@entry=0x7ffca2ba7320, format=format@entry=0x7fe0dec9bb1e ", %s", ap=ap@entry=0x7ffca2ba7500) at vfprintf.c:1643
      #2  0x00007fe0e0ed3910 in _IO_vsnprintf (string=0x7ffca2ba76c7 ", ", maxlen=<optimised out>, format=0x7fe0dec9bb1e ", %s", args=0x7ffca2ba7500) at vsnprintf.c:114
      #3  0x00007fe0debe2084 in __wt_vsnprintf_len_incr (buf=0x7ffca2ba76c7 ", ", size=3977, retsizep=0x7ffca2ba7638, fmt=0x7fe0dec9bb1e ", %s", ap=0x7ffca2ba7500) at ../src/os_posix/os_snprintf.c:21
      #4  0x00007fe0dec37e71 in __wt_snprintf_len_set (buf=0x7ffca2ba76c7 ", ", size=3977, retsizep=0x7ffca2ba7638, fmt=0x7fe0dec9bb1e ", %s") at ../src/include/misc.i:110
      #5  0x00007fe0dec38819 in __eventv (session=0x5582677384a0, msg_event=false, error=0, func=0x7fe0dec86580 <__func__.35607> "__curfile_close", line=511, fmt=0x7fe0dec85b68 "%s", ap=0x7ffca2ba8770) at ../src/support/err.c:219
      #6  0x00007fe0dec38f56 in __wt_errx_func (session=0x5582677384a0, func=0x7fe0dec86580 <__func__.35607> "__curfile_close", line=511, fmt=0x7fe0dec85b68 "%s") at ../src/support/err.c:330
      #7  0x00007fe0deb57067 in __curfile_close (cursor=0x55826778eba0) at ../src/cursor/cur_file.c:511
      #8  0x00007fe0dec1f53f in __session_close_cursors (session=0x5582677384a0, cursors=0x558267761350) at ../src/session/session_api.c:234
      #9  0x00007fe0dec1f5de in __session_close_cached_cursors (session=0x5582677384a0) at ../src/session/session_api.c:252
      #10 0x00007fe0dec1fa5a in __session_close (wt_session=0x5582677384a0, config=0x0) at ../src/session/session_api.c:288
      #11 0x00007fe0deb363a4 in __wt_connection_close (conn=0x55826771e010) at ../src/conn/conn_open.c:152
      #12 0x00007fe0deb2276a in __conn_close (wt_conn=0x55826771e010, config=0x0) at ../src/conn/conn_api.c:1100
      

      Note: As part of the patch, I have disabled cursor caching for history store cursors:

      $ cat diff 
      diff --git a/src/cursor/cur_file.c b/src/cursor/cur_file.c
      index 32adec887..cce07a661 100644
      --- a/src/cursor/cur_file.c
      +++ b/src/cursor/cur_file.c
      @@ -720,8 +720,11 @@ __curfile_create(WT_SESSION_IMPL *session, WT_CURSOR *owner, const char *cfg[],
       
           /*
            * WiredTiger.wt should not be cached, doing so interferes with named checkpoints.
      +     * WiredTigerHS.wt should not be cached <-- Needs a discussion whether we should cache HS
      +     * cursors.
            */
      -    if (cacheable && strcmp(WT_METAFILE_URI, cursor->internal_uri) != 0)
      +    if (cacheable && strcmp(WT_METAFILE_URI, cursor->internal_uri) != 0 &&
      +      strcmp(WT_HS_URI, cursor->internal_uri) != 0)
               F_SET(cursor, WT_CURSTD_CACHEABLE);
       
           WT_ERR(__wt_cursor_init(cursor, cursor->internal_uri, owner, cfg, cursorp));
      

      This ticket is to understand how to fix cursor caching for history store cursor. Before digging too much into a "fix", also to keep in mind:

      • If we need cursor caching for history store cursors or an alternate mechanism would be more efficient. Neither we want to open/close history store cursor each time required, nor we want to build an alternative mechanism to do what cursor caching already achieves.

        1. patch
          77 kB
          Sulabh Mahajan

            Assignee:
            haseeb.bokhari@mongodb.com Haseeb Bokhari (Inactive)
            Reporter:
            sulabh.mahajan@mongodb.com Sulabh Mahajan
            Votes:
            0 Vote for this issue
            Watchers:
            3 Start watching this issue

              Created:
              Updated:
              Resolved: