unflatten() decides in a first pass which prefixes of the keys denote arrays
(those that have a reference token 0 among their children), and looks up
prefixes again in the second pass. A prefix was identified by hashing it, and
by comparing it byte by byte with the entry in the set, from its beginning.
A key with n tokens therefore cost n^2/2 bytes: 100,000 tokens "/0" took
2.4 s (3.12.0: 5 ms), and so did "/0" followed by "/1" tokens, because every
new level looks its prefix up. Keys in the order of an unsorted object type
(ordered_json) defeated the shortcut for keys that share a beginning with the
previous one: four keys were enough to be quadratic.
Prefixes of less than 64 bytes are kept in a hash table and compared as
before, which costs a bounded number of bytes. Longer prefixes are kept in a
tree whose nodes are looked up by their parent and the text of one token. The
node of a prefix is found by walking the tokens of the key from the root, so
all prefixes of a key together cost one pass over the key, and no prefix is
compared as a whole. The nodes along the key walked last are kept, so keys
that share a beginning do not walk it again. The array/object decision and
the independence from the order of the keys are unchanged.
The hash function is seeded with a value derived from the address of a local
variable, which address space layout randomization changes between
processes, so keys cannot be crafted to collide in advance: with the fixed
constants, 20,000 keys of 16 + 2 bytes with the same hash took 0.5 s (random
keys, and 3.12.0: 7 ms). Where addresses are not randomized, or when the seed
is known, k short prefixes can still collide and cost k^2 / 2 comparisons of
fewer than 64 bytes. Results do not depend on the seed.
Measurements (Apple clang -O2, arm64), unflatten() of one key with n tokens:
tokens "/0" n=25k n=50k n=100k n=200k
before 175 ms 617 ms 2437 ms 9729 ms
after 4 ms 6 ms 12 ms 27 ms
3.12.0 1 ms 3 ms 5 ms 10 ms
"/0" then "/1" n=25k n=50k n=100k n=200k
before 153 ms 603 ms 2436 ms 9678 ms
after 1 ms 3 ms 5 ms 10 ms
3.12.0 1 ms 3 ms 5 ms 11 ms
ordered_json, 4 keys /0*m/a0, /10, /0*m/a1, /11
m=25k m=50k m=100k m=200k
before 397 ms 1402 ms 5191 ms 20723 ms
after 4 ms 7 ms 15 ms 33 ms
3.12.0 2 ms 3 ms 8 ms 17 ms
Ordinary documents (flatten() then unflatten(), instructions retired): within
0.3 % of before for canada, citm_catalog, twitter, jeopardy, mesh,
update-center and random from the benchmark corpus.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann
b3dd34036a