IValue: environment friendly illustration of dynamic sorts in C++

Introduction

In conventional SQL programs, a column’s sort is decided when the desk is created, and by no means modifications whereas executing a question. If you happen to create a desk with an integer-valued column, the values in that column will all the time be integers (or probably NULL).

Rockset, nevertheless, is dynamically typed, which signifies that we regularly do not know the kind of a worth till we really execute the question. That is just like different dynamically typed programming languages, the place the identical variable could include values of various sorts at completely different closing dates:

$ python3
>>> a = 3
>>> sort(a)
<class 'int'>
>>> a="foo"
>>> sort(a)
<class 'str'>

Rockset’s sort system was initially primarily based on JSON, and has since been prolonged to help different sorts as effectively:

• bytes: taking a cue from Python, we distinguish between sequences of legitimate Unicode characters (string, which is internally represented as UTF-8) and sequences of arbitrary bytes (bytes)
• date- and time-specific sorts (date, time, datetime, timestamp, microsecond_interval, month_interval)

There are different sorts that we use internally (and are by no means uncovered to our customers); additionally, the sort system is extensible, with deliberate help for decimal (base-10 floating-point), geometry / geography sorts, and others.

Within the following instance, assortment ivtest has paperwork containing one discipline a, which takes a wide range of sorts:

$ rock create assortment ivtest
Assortment "ivtest" was created efficiently in workspace "commons".

$ cat /tmp/a.docs
{"a": 2}
{"a": "good day"}
{"a": null}
{"a": {"b": 10}}
{"a": [2, "foo"]}

$ rock add ivtest /tmp/a.docs
{
 "file_name":"a.docs",
 "file_upload_id":"c5ccc261-0096-4a73-8dfe-d6db8b8d130e",
 "uploaded_at":"2019-06-05T18:12:46Z"
}

$ rock sql
> choose typeof(a), a from ivtest order by a;
+-----------+------------+
| ?typeof   | a          |
|-----------+------------|
| null_type | <null>     |
| int       | 2          |
| string    | good day      |
| array     | [2, 'foo'] |
| object    | {'b': 10}  |
+-----------+------------+
Time: 0.014s

This put up reveals considered one of many challenges that we encountered whereas constructing a completely dynamically typed SQL database: how we manipulate values of unknown sorts in our question execution backend (written in C++), whereas approaching the efficiency of utilizing native sorts straight.

At first, we used protocol buffers just like the definition beneath (simplified to solely present integers, floats, strings, arrays, and objects; the precise oneof that we use has a couple of further fields):

message Worth {
  oneof value_union {
    int64 int_value = 1;
    double float_value = 2;
    string string_value = 3;
    ArrayValue array_value = 4;
    ObjectValue object_value = 5;
  }
}

message ArrayValue {
  repeated Worth values = 1;
}

message ObjectValue {
  repeated KeyValue kvs = 1;
}

message KeyValue {
  string key = 1;
  Worth worth = 2;
}   

However we rapidly realized that that is inefficient, each by way of velocity and by way of reminiscence utilization. First, protobuf requires a heap reminiscence allocation for each object; making a Worth that comprises an array of 10 integers would carry out:

• a reminiscence allocation for the top-level Worth
• an allocation for the array_value member
• an allocation for the record of values (ArrayValue.values, which is a RepeatedPtrField)
• an allocation for every of the ten values within the array

for a complete of 13 reminiscence allocations.

Additionally, the ten values within the array aren’t allotted contiguously in reminiscence, which causes an additional lower in efficiency because of cache locality.

It was rapidly clear that we would have liked one thing higher, which we known as IValue. In comparison with the protobuf model, IValue is:

• Extra reminiscence environment friendly: whereas not as environment friendly as utilizing native sorts straight, IValue have to be small, and should keep away from heap allocations wherever doable. IValue is all the time 16 bytes, and doesn’t allocate heap reminiscence for integers, booleans, floating-point numbers, and quick strings.
• Quicker: arrays of scalar IValues are allotted contiguously in reminiscence, main to raised cache locality. This isn’t as environment friendly as utilizing native sorts straight, however it’s a vital enchancment over protobuf.

Most of Rockset’s question execution engine operates on IValues (there are some components which have specialised implementation for particular sorts, and that is an space of energetic enchancment).

We might wish to share an summary of the IValue design. Be aware that IValue is optimized for Rockset’s wants and isn’t meant to be transportable — we use Linux and x86_64-specific methods, and assume a little-endian reminiscence structure.

The concept is in itself not novel; the methods that we use date again to no less than 1993, as surveyed in “Representing Sort Data in Dynamically Typed Languages”. We determined to make IValue 128 bits as a substitute of 64, because it permits us to keep away from heap allocations in additional circumstances (together with all 64-bit integers); utilizing the taxonomy outlined within the paper, IValue is a double-wrapper scheme with qualifiers.

Internally, IValue is represented as a 128-bit (16-byte) worth, consisting of:

• a 64-bit discipline (known as knowledge)
• a 48-bit discipline (known as pointer, because it usually, however not all the time, shops a pointer)
• two 8-bit discriminator fields (known as tag0 and tag1)

tag1 signifies the kind of the worth. tag0 is often a subtype, and the that means of the opposite two fields modifications relying on sort. The pointer discipline is usually a pointer to another knowledge construction, allotted on the heap, for the circumstances the place heap allocations cannot be averted; as pointers are solely 48 bits on x86_64, we’re capable of match a pointer and the 2 discriminator fields in the identical uint64_t.

We acknowledge two kinds of IValues:

tag1 has bit 7 clear (tag1 < 0x80) for all instant values, and set (tag1 >= 0x80) for all non-immediate values. This enables us to differentiate between instant and non-immediate values in a short time, utilizing one easy bit operation. We will then copy, hash, and examine for equality instant values by treating them as a pair of uint64_t integers.

Scalar Varieties

The illustration for many scalar sorts is simple: tag0 is often zero, tag1 identifies the sort, pointer is often zero, and knowledge comprises the worth.

SQL NULL is all zeros, which is handy (memset()ing a piece of reminiscence to zero makes it NULL when interpreted as IValue):

Booleans have knowledge = 0 for false and knowledge = 1 for true, tag1 = 0x01

Integers have the worth saved in knowledge (as int64_t) and tag1 = 0x02

And so forth. The layouts for different scalar sorts (floating level, date / time, and many others) are related.

Strings

We deal with character strings and byte strings equally; the worth of tag1 is the one distinction. For the remainder of the part, we’ll solely deal with character strings.

IValue strings are immutable, preserve the string’s size explicitly, and aren’t null-terminated. In step with our aim to attenuate heap allocations, IValue would not use any exterior reminiscence for brief strings (lower than 16 bytes).

As a substitute, we implement the small string optimization: we retailer the string contents (padded with nulls) within the knowledge, pointer, and tag0 fields; we retailer the string size within the tag1 discipline: tag1 is 0x1n, the place n is the string’s size.

An empty string has tag1 = 0x10 and all different bytes zero:

And, for instance, the 11-byte string “Hi there world” has tag1 = 0x1b (be aware the little-endian illustration; the byte 'H' is first):

Strings longer than 15 bytes are saved out-of-line: tag1 is 0x80, pointer factors to the start of the string (allotted on the heap utilizing malloc()), and knowledge comprises the string size. (There’s additionally the potential for referencing a “overseas” string, the place IValue would not personal the reminiscence however factors inside a preallocated buffer, however that’s past the scope of this put up.)

For instance, the 19-byte string “Rockset is superior!”:

Vectors

Vectors (which we name “arrays”, adopting JSON’s terminology) are equally allotted on the heap: they’re just like vectors in most programming languages (together with C++’s std::vector). tag1 is 0x82, pointer factors to the start of the vector (allotted on the heap utilizing malloc()), and knowledge comprises the vector’s dimension and capability (32 bits every). The vector itself is a contiguously allotted block of capability() IValues (capability() * 16 bytes); when reallocation is required, the vector grows exponentially (with an element that’s lower than 2, for the explanations described in Fb’s fbvector implementation.)

Hash Maps

Maps (which we name “objects”, adopting JSON’s terminology) are additionally allotted on the heap. We symbolize objects as open-addressing hash tables with quadratic probing; the dimensions of the desk is all the time an influence of two, which simplifies probing. We probe with triangular numbers, identical to Google’s sparsehash, which. as Knuth tells us in The Artwork of Pc Programming (quantity 3, chapter 6.4, train 20), mechanically covers all slots.

Every hash desk slot is 32 bytes — two IValues, one for the important thing, one for the worth. As is often the case with open-addressing hash tables, we’d like two particular keys — one to symbolize empty slots, and one to symbolize deleted parts (tombstones). We reserve two values of tag1 for that objective (0x06 and 0x05, respectively).

The pointer discipline factors to the start of the hash desk (a contiguous array of slots, allotted on the heap utilizing malloc().) We retailer the present dimension of the hash desk within the least-significant 32 bits of the knowledge discipline. The tag0 discipline comprises the variety of allotted slots (because it’s all the time an influence of two, we retailer log2(variety of slots) + 1, or zero if the desk is empty).

The capability discipline (most important 32 bits of knowledge) deserves additional curiosity: it’s the variety of slots obtainable for storing person knowledge. Initially, it’s the similar as the entire variety of slots, however, as in all open-addressing hash tables, erasing a component from the desk marks the slot as “deleted” and renders it unusable till the subsequent rehash. So erasing a component really decreases the desk’s capability.

Efficiency

IValue provides a considerable efficiency enchancment over the previous protobuf-based implementation:

• creating arrays of strings is between 2x and 7x sooner (relying on the string dimension; due to the small-string optimization, IValue is considerably sooner for small strings)
• creating arrays of integers can be 7x sooner (as a result of we not allocate reminiscence for each particular person array ingredient)
• iterating over massive arrays of integers is 3x sooner (as a result of the values within the array at the moment are allotted contiguously)

Future Work

Though Rockset paperwork are allowed to include knowledge of a number of sorts in the identical discipline, the state of affairs proven within the introduction is comparatively uncommon. In follow, many of the knowledge is of the identical sort (or NULL), and, to acknowledge this, we’re extending IValue to help homogeneous arrays.

All parts in a homogeneous array are of the identical sort (or NULL). The construction is just like the common (heterogeneous) arrays (described above), however the pointer discipline factors on to an array of the native sort (int64_t for an array of integers, double for an array of floating-point values, and many others). Just like programs like Apache Arrow, we additionally preserve an non-obligatory bitmap that signifies whether or not a selected worth is NULL or not.

The question execution code acknowledges the frequent case the place it produces a column of values of the identical sort, wherein case it’s going to generate a homogeneous array. We now have environment friendly, vectorized implementations of frequent database operations on homogeneous arrays, permitting us vital efficiency enhancements within the frequent case.

That is nonetheless an space of energetic work, and benchmark outcomes are forthcoming.

Conclusion

We hope that you just loved a quick look beneath the hood of Rockset’s engine. Sooner or later, we’ll share extra particulars about our approaches to constructing a completely dynamically typed SQL database; if you would like to provide us a attempt, join an account; if you would like to assist construct this, we’re hiring!