5.2 KiB
HogVM
A HogVM is a 🦔 that runs HogQL bytecode. It's purpose is to locally evaluate HogQL expressions against any object.
HogQL bytecode
HogQL Bytecode is a compact representation of a subset of the HogQL AST nodes. It follows a certain structure:
1 + 2 # [_H, op.INTEGER, 2, op.INTEGER, 1, op.PLUS]
1 and 2 # [_H, op.INTEGER, 2, op.INTEGER, 1, op.AND, 2]
1 or 2 # [_H, op.INTEGER, 2, op.INTEGER, 1, op.OR, 2]
not true # [_H, op.TRUE, op.NOT]
properties.bla # [_H, op.STRING, "bla", op.STRING, "properties", op.FIELD, 2]
call('arg', 'another') # [_H, op.STRING, "another", op.STRING, "arg", op.CALL, "call", 2]
1 = 2 # [_H, op.INTEGER, 2, op.INTEGER, 1, op.EQ]
'bla' !~ 'a' # [_H, op.STRING, 'a', op.STRING, 'bla', op.NOT_REGEX]
Compliant implementation
The python/execute.py
function in this folder acts as the reference implementation in case of disputes.
Operations
To be considered a PostHog HogQL Bytecode Certified Parser, you must implement the following operations:
FIELD = 1 # [arg3, arg2, arg1, FIELD, 3] # arg1.arg2.arg3
CALL = 2 # [arg2, arg1, CALL, 'concat', 2] # concat(arg1, arg2)
AND = 3 # [val3, val2, val1, AND, 3] # val1 and val2 and val3
OR = 4 # [val3, val2, val1, OR, 3] # val1 or val2 or val3
NOT = 5 # [val, NOT] # not val
PLUS = 6 # [val2, val1, PLUS] # val1 + val2
MINUS = 7 # [val2, val1, MINUS] # val1 - val2
MULTIPLY = 8 # [val2, val1, MULTIPLY] # val1 * val2
DIVIDE = 9 # [val2, val1, DIVIDE] # val1 / val2
MOD = 10 # [val2, val1, MOD] # val1 % val2
EQ = 11 # [val2, val1, EQ] # val1 == val2
NOT_EQ = 12 # [val2, val1, NOT_EQ] # val1 != val2
GT = 13 # [val2, val1, GT] # val1 > val2
GT_EQ = 14 # [val2, val1, GT_EQ] # val1 >= val2
LT = 15 # [val2, val1, LT] # val1 < val2
LT_EQ = 16 # [val2, val1, LT_EQ] # val1 <= val2
LIKE = 17 # [val2, val1, LIKE] # val1 like val2
ILIKE = 18 # [val2, val1, ILIKE] # val1 ilike val2
NOT_LIKE = 19 # [val2, val1, NOT_LIKE] # val1 not like val2
NOT_ILIKE = 20 # [val2, val1, NOT_ILIKE] # val1 not ilike val2
IN = 21 # [val2, val1, IN] # val1 in val2
NOT_IN = 22 # [val2, val1, NOT_IN] # val1 not in val2
REGEX = 23 # [val2, val1, REGEX] # val1 =~ val2
NOT_REGEX = 24 # [val2, val1, NOT_REGEX] # val1 !~ val2
IREGEX = 25 # [val2, val1, IREGEX] # val1 =~* val2
NOT_IREGEX = 26 # [val2, val1, NOT_IREGEX] # val1 !~* val2
TRUE = 29 # [TRUE] # true
FALSE = 30 # [FALSE] # false
NULL = 31 # [NULL] # null
STRING = 32 # [STRING, 'text'] # 'text'
INTEGER = 33 # [INTEGER, 123] # 123
FLOAT = 34 # [FLOAT, 123.12] # 123.01
Async Operations
Some operations can't be computed directly, and are thus asked back to the caller. These include:
IN_COHORT = 27 # [val2, val1, IREGEX] # val1 in cohort val2
NOT_IN_COHORT = 28 # [val2, val1, NOT_IREGEX] # val1 not in cohort val2
The arguments for these instructions will be passed on to the provided async_operation(*args)
in reverse:
def async_operation(*args):
if args[0] == op.IN_COHORT:
return db.queryInCohort(args[1], args[2])
return False
execute_bytecode(to_bytecode("'user_id' in cohort 2"), {}, async_operation).result
Functions
A PostHog HogQL Bytecode Certified Parser must also implement the following function calls:
concat(...) # concat('test: ', 1, null, '!') == 'test: 1!'
match(string, pattern) # match('fish', '$fi.*') == true
toString(val) # toString(true) == 'true'
toInt(val) # toInt('123') == 123
toFloat(val) # toFloat('123.2') == 123.2
toUUID(val) # toUUID('string') == 'string'
ifNull(val, alternative) # ifNull('string', false) == 'string'
Null handling
In HogQL equality comparisons, null
is treated as any other variable. Its presence will not make functions automatically return null
, as is the ClickHouse default.
1 == null # false
1 != null # true
Nulls are just ignored in concat
Known broken features
- Regular Expression support is implemented, but NOT GUARANTEED to the same way across platforms. Different implementations (ClickHouse, Python, Node) use different Regexp engines. ClickHouse uses
re2
, the others usepcre
. Use the case-insensitive regex operators instead of passing in modifier flags through the expression. - DateTime comparisons are not supported.
- Cohort Matching operations are not implemented.
- Only a small subset of functions is enabled. This list is bound to expand.