fixup! ZBigFile: Add ZBlk format option 'h' (heuristic) (4)

Try to keep documented promise that

	The heuristic 'auto' should behave as good as ZBlk0 in case of wide changes

Previously the heuristic code would use ZBlk0 only in append case after
switching previously filled-up block to ZBlk0, and the case of arbitrary
changes were unconditionally using ZBlk1.

Now we see if the append is small and use ZBlk1 only then. If the append
is big, or if the change is itself is big - we use ZBlk0.

This should restore documented behaviour that heuristic behaves as good
as ZBlk0 in case of wide changes.

TODO update benchmark/test to cover this case.

bigfile/file_zodb.py

@@ -587,20 +587,38 @@ class ZBigFile(LivePersistent):
     # servers, wendelin.core cannot provide at the same time both
     # fast reads and small database size growth ..."
     def _zblk_fmt_heuristic(self, zblk, blk, buf):
-        if _is_appending(zblk, buf):
+        # see if we are doing a "small append" like change
+        # load previous data along the way
+        new_data = bytes(buf).rstrip(b'\0')
+        old_data = b''
+        if zblk is None:
+            appending = (len(new_data) < 0.5*self.blksize)
+        else:
+            old_data  = bytes(zblk.loadblkdata()).rstrip(b'\0')
+            appending = new_data[:len(old_data)] == old_data
+
+        # append - use ZBlk1 and migrate previously filled-up block to ZBlk0 for fast reads
+        if appending:
             if not zblk and blk > 0:  # is new zblk?
-                # Set previous filled-up ZBlk to ZBlk0 for fast reads
-                previous_blk = blk - 1
-                previous_zblk = self.blktab.get(previous_blk)
-                if previous_zblk is not None:
-                    self._setzblk(previous_blk, previous_zblk, previous_zblk.loadblkdata(), ZBlk0)
+                zblk_prev = self.blktab.get(blk-1)
+                if zblk_prev is not None  and  type(zblk_prev) is not ZBlk0:
+                    self._setzblk(blk-1, zblk_prev, zblk_prev.loadblkdata(), ZBlk0)
             return ZBlk1
-        else:  # it's changing
-            # kirr: "to support sporadic small changes over initial big fillup [...]
+
+        # arbitrary change - use ZBlk1 if the change is small and ZBlk0 otherwise
+        else:
+            # TODO(kirr): "to support sporadic small changes over initial big fillup [...]
             # we could introduce e.g. a ZBlkδ object, which would refer to base
             # underlying ZBlk object and add "patch" information on top of that [...]."
             # See https://lab.nexedi.com/nexedi/wendelin.core/merge_requests/20#note_196084
-            return ZBlk1
+            ndelta = 0
+            for i in range(self.blksize):
+                if new_data[i:i+1] != old_data[i:i+1]:
+                    ndelta += 1
+            if ndelta < 0.5*self.blksize:
+                return ZBlk1
+            else:
+                return ZBlk0
 
 
     # invalidate data   .blktab[blk] invalidated -> invalidate page
@@ -865,11 +883,3 @@ class _ZBigFileH(object):
         # and also more right - tpc_finish is there assumed as non-failing by
         # ZODB design)
         self.abort(txn)
-
-
-# Utility functions for heuristic
-def _is_appending(zblk, buf):
-    if not zblk:
-        return True
-    old_buf = bytes(zblk.loadblkdata())
-    return bytes(buf).rstrip(b'\0')[:len(old_buf)] == old_buf