Regex search is extremely slow with python3

Hi.

As is described here:

in Python 3, a regular expression compiled from a str has the re.UNICODE flag set.

But unicode regexps is extremely slow:

2015-04-18T23:44:29.539+0300 I QUERY    [conn2] query test.computers query: { $or: [ { number: "test" }, { hostname: /^test/u }, { macs: /^test/u }, { ipmi: /^test/u }, { dc: /^test/u } ] } planSummary: IXSCAN { hostname: 1 }, IXSCAN { number: 1 }, IXSCAN { dc: 1 }, IXSCAN { macs: 1 }, IXSCAN { ipmi: 1 } ntoreturn:100 ntoskip:0 nscanned:498422 nscannedObjects:104692 keyUpdates:0 writeConflicts:0 numYields:4391 nreturned:3 reslen:643 locks:{} 1001ms

Non-unicode version:

2015-04-18T23:42:33.177+0300 I QUERY    [conn1] query test.computers query: { $or: [ { number: "test" }, { hostname: /^test/ }, { macs: /^test/ }, { ipmi: /^test/ }, { dc: /^test/ } ] } planSummary: IXSCAN { hostname: 1 }, IXSCAN { number: 1 }, IXSCAN { dc: 1 }, IXSCAN { macs: 1 }, IXSCAN { ipmi: 1 } ntoreturn:100 ntoskip:0 nscanned:4 nscannedObjects:3 keyUpdates:0 writeConflicts:0 numYields:0 nreturned:3 reslen:643 locks:{} 1ms

I've tried to use such method in my model:

@staticmethod
def search(pattern):
    return Computer.objects.filter(
        db.Q(number=pattern) |
        db.Q(hostname__startswith=pattern) |
        db.Q(macs__startswith=pattern) |
        db.Q(ipmi__startswith=pattern) |
        db.Q(dc__startswith=pattern)
    )

But it generates unicode regex (1st query).
Now I'm using different method:

@staticmethod
def search(pattern):
  r = bson.regex.Regex('^{}'.format(escape(pattern)))
  return Computer.objects(
      db.Q(number=pattern) |
      db.Q(hostname=r) |
      db.Q(macs=r) |
      db.Q(ipmi=r) |
      db.Q(dc=r)
  )

And it works fine (2nd query). But it is terrible, I think.
Also I've tried to use bytestring instead of unicode string, but it handles invalid here.

Read mongoengine/fields.py lines 90-111, then reproduce the two search forms from the issue under Python 3. Compare the generated MongoDB regex options and query performance, including the bytestring case. Done should mean the documented startswith query no longer produces the unexpectedly slow unicode-regex behavior, or the limitation is clearly documented.