sparse-read: move from a recursive-based approach to a heap-based one

The previous recursive approach was trying to optimise each read slice to have
a good density. It had the tendency to over-optimize smaller slices while
leaving larger hole in others.

The new approach focuses on improving the combined density of all the reads,
instead of the individual slices. It slices at the largest gaps first, as they
reduce the total amount of read data the most efficiently.

Another benefit of this approach is that we iterate over the delta chain only
once, reducing the overhead of slicing long delta chains.

On the repository we use for tests, the new approach shows similar or faster
performance than the current default linear full read.

The repository contains about 450,000 revisions with many concurrent
topological branches. Tests have been run on two versions of the repository:
one built with the current delta constraint, and the other with an unlimited
delta span (using 'experimental.maxdeltachainspan=0')

Below are timings for building 1% of all the revision in the manifest log using
'hg perfrevlogrevisions -m'. Times are given in seconds. They include the new
couple of follow-up changeset in this series.

delta-span  standard    unlimited
linear-read     922s         632s
sparse-read     814s         566s

mercurial/revlog.py

@@ -17,6 +17,7 b' import binascii'
 import collections
 import errno
 import hashlib
+import heapq
 import os
 import struct
 import zlib
@@ -170,49 +171,59 b' def _slicechunk(revlog, revs):'
     start = revlog.start
     length = revlog.length
 
-    chunkqueue = collections.deque()
-    chunkqueue.append((revs, 0))
-
-    while chunkqueue:
-        revs, depth = chunkqueue.popleft()
+    if len(revs) <= 1:
+        yield revs
+        return
 
         startbyte = start(revs[0])
         endbyte = start(revs[-1]) + length(revs[-1])
-        deltachainspan = endbyte - startbyte
+    readdata = deltachainspan = endbyte - startbyte
 
-        if deltachainspan <= revlog._srminblocksize or len(revs) <= 1:
-            yield revs
-            continue
+    chainpayload = sum(length(r) for r in revs)
 
-        # Find where is the largest hole (this is where we would split) and
-        # sum up the lengths of useful data to compute the density of the span
-        textlen = 0
+    if deltachainspan:
+        density = chainpayload / float(deltachainspan)
+    else:
+        density = 1.0
+
+    # Store the gaps in a heap to have them sorted by decreasing size
+    gapsheap = []
+    heapq.heapify(gapsheap)
         prevend = None
-        largesthole = 0
-        idxlargesthole = -1
         for i, rev in enumerate(revs):
             revstart = start(rev)
             revlen = length(rev)
 
             if prevend is not None:
-                hole = revstart - prevend
-                if hole > largesthole:
-                    largesthole = hole
-                    idxlargesthole = i
+            gapsize = revstart - prevend
+            if gapsize:
+                heapq.heappush(gapsheap, (-gapsize, i))
 
-            textlen += revlen
             prevend = revstart + revlen
 
-        density = textlen / float(deltachainspan) if deltachainspan > 0 else 1.0
+    # Collect the indices of the largest holes until the density is acceptable
+    indicesheap = []
+    heapq.heapify(indicesheap)
+    while gapsheap and density < revlog._srdensitythreshold:
+        oppgapsize, gapidx = heapq.heappop(gapsheap)
+
+        heapq.heappush(indicesheap, gapidx)
 
-        if density > revlog._srdensitythreshold:
-            yield revs
-            continue
+        # the gap sizes are stored as negatives to be sorted decreasingly
+        # by the heap
+        readdata -= (-oppgapsize)
+        if readdata > 0:
+            density = chainpayload / float(readdata)
+        else:
+            density = 1.0
 
-        # Add the left and right parts so that they will be sliced
-        # recursively too
-        chunkqueue.append((revs[:idxlargesthole], depth + 1))
-        chunkqueue.append((revs[idxlargesthole:], depth + 1))
+    # Cut the revs at collected indices
+    previdx = 0
+    while indicesheap:
+        idx = heapq.heappop(indicesheap)
+        yield revs[previdx:idx]
+        previdx = idx
+    yield revs[previdx:]
 
 # index v0:
 #  4 bytes: offset