Redis int representation of a string is bigger when the string is more than 7 bytes but smaller otherwise - optimization

Question

I'm trying to reduce Redis's objects size as much as I can and I've taken this whole week to experiment with it.
While testing different data representations I found out that an int representation of the string "hello" results in a smaller object. It may not look like much, but if you have a lot of data it can make a difference between using a few GB memory vs dozens of it.
Look at the following example (you can try it yourself if you want):
> SET test:1 "hello"
> debug object test:1
> Value at:0xb6c9f380 refcount:1 encoding:raw serializedlength:6 lru:9535350 lru_seconds_idle:7
In particular you should look at serializedlength which is 6 (bytes) in this case.
Now, look at the following int representation of it:
> SET test:2 "857715"
> debug object test:2
> Value at:0xb6c9f460 refcount:1 encoding:int serializedlength:5 lru:9535401 lru_seconds_idle:2
As you see, it results in a byte shorter object (note also encoding:int which I think is suggesting that ints get handled in a more efficient way).
With the string "hello w" (you'll see in a few moments why I didn't use "hello world" instead) we get an even bigger saving when it's represented as an int:
> SET test:3 "hello w"
> SET test:4 "857715023" <- Int representation. Notice that I inserted a "0", if I don't, it results in a bigger object and the encoding is set to "raw" instead (after all a space is not an int).
>
> debug object test:3
> Value at:0xb6c9f3a0 refcount:1 encoding:raw serializedlength:8 lru:9535788 lru_seconds_idle:6
> debug object test:4
> Value at:0xb6c9f380 refcount:1 encoding:int serializedlength:5 lru:9535809 lru_seconds_idle:5
It looks cool as long as you don't exceed 7 bytes string.. Look at what happens by a "hello wo" int representation:
> SET test:5 "hello wo"
> SET test:6 "85771502315"
>
> debug object test:5
> Value at:0xb6c9f430 refcount:1 encoding:raw serializedlength:9 lru:9535907 lru_seconds_idle:9
> debug object test:6
> Value at:0xb6c9f470 refcount:1 encoding:raw serializedlength:12 lru:9535913 lru_seconds_idle:5
As you can see the int (12 bytes) is bigger than the string representation (9 bytes).
My question here is, what's going on behind the scenes when you represent a string as an int, that it is smaller until you reach 7 bytes?
Is there a way to increase this limit as you do with "list-max-ziplist-entries/list-max-ziplist-value" or a clever way to optimize this process so that it always (or nearly) results in a smaller object than a string?
UPDATE
I've further experimented with other tricks, and you can actually have smaller ints than string, regardless of its size, but that would involve a little more work as of data structure modelling.
I've found out that if you split the int representation of a string in chunks of ~8 numbers each, it ends up being smaller.
Take as an example the word "Hello World Hi Universe" and create both a string and int SET:
> HMSET test:7 "Hello" "World" "Hi" "Universe"
> HMSET test:8 "74111114" "221417113" "78" "2013821417184"
The results are as follows:
> debug object test:7
> Value at:0x7d12d600 refcount:1 encoding:ziplist serializedlength:40 lru:9567096 lru_seconds_idle:296
>
> debug object test:8
> Value at:0x7c17d240 refcount:1 encoding:ziplist serializedlength:37 lru:9567531 lru_seconds_idle:2
As you can see we got the int set smaller by 3 bytes.
The problem in this will be how to organize such a thing, but it shows that it's possible nonetheless.
Still, don't know where this limit is set. The ~700K persistent use of memory (even when you have no data inside) makes me think that there is a pre-defined "pool" dedicated to the optimization of int sets.
UPDATE2
I think I've found where this intset "pool" is defined in Redis source.
At line 81 in the file redis.h there is the def REDIS_SHARED_INTEGERS set to 10000
REDISH_SHARED_INTEGERS
I suspect it's the one defining the limit of an intset byte length.
I have to try to recompile it with an higher value and see if I can use a longer int value (it'll most probably allocate more memory if it's the one I think of).
UPDATE3
I want to thank Antirez for the reply! Didn't expect that.
As he made me notice, len != memory usage.
I got further in my experiment and saw that the objects get already slightly compressed (serialized). I may have missed something from the Redis documentation.
The confirmation comes from analyzing a Redis key wih the command redis-memory-for-key key, which actually returns the memory usage and not the serialized length.
For example, let's take the "hello" string and int we used before, and see what's the result:
~ # redis-memory-for-key test:1
Key "test:1"
Bytes 101
Type string
~ #
~ # redis-memory-for-key test:2
Key "test:2"
Bytes 87
Type string
As you can notice the intset is smaller (87 bytes) than the string (101 bytes) anyway.
UPDATE4
Surprisingly a longer intset seems to affect its serializedlength but not memory usage..
This makes it possible to actually build a 2digit-char mapping while it still being more memory efficient than a string, without even chunking it.
By 2digit-char mapping I mean that instead of mapping "hello" to "85121215" we map it to digits with a fixed length of 2 each, prefixing it with "0" if digit < 10 like "0805121215".
A custom script would then proceed by taking every two digit apart and converting them to their equivalent char:
08 05 12 12 15
\ | | | /
h e l l o
This is enough to avoid disambiguation (like "o" and "ae" which both result in the digit "15").
I'll show you this works by creating another set and therefore analyzing its memory usage like I did before:
> SET test:9 "0805070715"
Unix shell
----------
~ # redis-memory-for-key test:9
Key "test:9"
Bytes 87
Type string
You can see that we have a memory win here.
The same "hello" string compressed with Smaz for comparison:
>>> smaz.compress('hello')
'\x10\x98\x06'
// test:10 would be unfair as it results in a byte longer object
SET post:1 "\x10\x98\x06"
~ # redis-memory-for-key post:1
Key "post:1"
Bytes 99
Type string

Answer 1

My question here is, what's going on behind the scenes when you represent a
string as an int, that it is smaller until you reach 7 bytes?
Notice that the integer you supplied as test #6 is no longer actually encoded
as an integer, but as raw:
SET test:6 "85771502315"
Value at:0xb6c9f470 refcount:1 encoding:raw serializedlength:12 lru:9535913 lru_seconds_idle:
So we see that a "raw" value occupies one byte plus the length of its string representation. In memory
you get that plus the overhead of the value.
The integer encoding, I suspect, encodes a number as a 32-bit integer; then it will always
need five bytes, one to tell its type, and four to store those 32 bits.
As soon as you overflow the maximum representable integer in 32 bits, which is either 2 billions or 4 depending on whether you use a sign or not, you need to revert to raw encoding.
So probably
2147483647 -> five bytes (TYPE_INT 0x7F 0xFF 0xFF 0xFF)
2147483649 -> eleven bytes (TYPE_RAW '2' '1' '4' '7' '4' '8' '3' '6' '4' '9')
Now, how can you squeeze a string representation PROVIDED THAT YOU ONLY USE AN ASCII SET?
You can get the string (140 characters):
When in the Course of human events it becomes necessary for one people
to dissolve the political bands which have connected them with another
and convert each character to a six-bit representation; basically its index in the string
"ABCDEFGHIJKLMNOPQRSTUVWXYZ01234 abcdefghijklmnopqrstuvwxyz56789."
which is the set of all the characters you can use.
You can now encode four such "text-only characters" in three "binary characters", a sort of "reverse base 64 encoding"; base64 encoding will get three binary characters and create a four-byte sequence of ASCII characters.
If we were to code it as groups of integers, we would save a few bytes - maybe get it down
to 130 bytes - at the cost of a larger overhead.
With this type of "reverse base64" encoding, we can get 140 character to 35 groups of four characters, which become a string of 35x3 = 105 binary characters, raw encoded to 106 bytes.
As long, I repeat, as you never use characters outside the range above. If you do, you can
enlarge the range to 128 characters and 7 bits, thus saving 12.5% instead of 25%; 140 characters will then become 126, raw encoded to 127 bytes, and you save (141-127) = 14 bytes.
Compression
If you have much longer strings, you can compress them (i.e., you use a function such as deflate() or gzencode() or gzcompress() ). Either straight; in which case the above string becomes 123 bytes. Easy to do.
Compressing many small strings: the Rube Goldberg approach
Since compression algorithms learn, and at the beginning they dare assume nothing, small strings will not compress greatly. They're "all beginning", so to speak. Just as an engine, when running cold the performances are inferior.
If you have a "corpus" of text these strings come from, you can use a time-consuming trick that "warms up" the compression engine and may double (or better) its performances.
Suppose you have two strings, COMMON and TARGET (the second one is the one you're interested in). If you z-compressed COMMON you would get, say, ZCMN. If you compressed TARGET you would get ZTRGT.
But as I said, since the gz compression algorithm is stream oriented, and it learns as it goes by, the compression ratio of the second half of any text (provided there aren't freakish statistical distribution changes between halves) is always appreciably higher than that of the first half.
So if you were to compress, say, COMMONTARGET, you'd get ZCMGHQI.
Notice that the first part of the string, as far as almost the end, is the same as before. Indeed if you compressed COMMONFOOBAR, you'd get something like ZCMQKL. And the second part is compressed better than before, even if we count the area of overlap as belonging entirely to the second string.
And this is the trick. Given a family of strings (TARGET, FOOBAR, CASTLE BRAVO), we compress not the strings, but the concatenation of those strings with a large prefix. Then we discard from the result the common compressed prefix. Thus TARGET is taken from the compression of COMMONTARGET (which is ZCMGHQI), and becomes GHQI instead of ZTRGT, with a 20% gain.
The decoder does the reverse: given GHQI, it first applies the common compressed prefix ZCM (which it must know); then it decodes the result, and finally discards the common uncompressed prefix, of which it need only know the length beforehand.
So the first sentence above (140 characters) becomes 123 when compressed by itself; if I take the rest of the Declaration and use it as a prefix, it compresses to 3355 bytes. This prefix plus my 140 bytes becomes 3409 bytes, of which 3352 are common, leaving 57 bytes.
At the cost of storing once the uncompressed prefix in the encoder, and the compressed prefix once in the decoder, and the whole thingamajig running five times as slow, I can now get those 140 bytes down to 57 instead of 123 - less than half of before.
This trick works great for small strings; for larger ones, the advantage isn't worth the pain. Also, different prefixes yield different results. The best prefixes are those that contain most of the sequences that are likely to appear in the string pool, ordered by increasing length.
Added bonus: the compressed prefix also doubles as a sort of weak encryption, as without that, you can't easily decode the compressed strings, even if you might be able to recover some pieces thereof.