add checksum offset to avoid bigint overflow

Author: vvkh

data_diff/databases/base.py

@@ -1156,6 +1156,17 @@ def is_autocommit(self) -> bool:
 _CHECKSUM_BITSIZE = CHECKSUM_HEXDIGITS << 2
 CHECKSUM_MASK = (2**_CHECKSUM_BITSIZE) - 1
 
+# bigint is typically 8 bytes
+# if checksum is shorter, most databases will pad it with zeros
+# 0xFF → 0x00000000000000FF;
+# because of that, the numeric representation is always positive,
+# which limits the number of checksums that we can add together before overflowing.
+# we can fix that by adding a negative offset of half the max value,
+# so that the distribution is from -0.5*max to +0.5*max.
+# then negative numbers can compensate for the positive ones allowing to add more checksums together
+# without overflowing.
+CHECKSUM_OFFSET = CHECKSUM_MASK // 2
+
 DEFAULT_DATETIME_PRECISION = 6
 DEFAULT_NUMERIC_PRECISION = 24
 

data_diff/databases/bigquery.py

@@ -36,6 +36,8 @@
     ConnectError,
     apply_query,
     QueryResult,
+    CHECKSUM_OFFSET,
+    CHECKSUM_HEXDIGITS
 )
 from data_diff.databases.base import TIMESTAMP_PRECISION_POS, ThreadLocalInterpreter, Mixin_RandomSample
 
@@ -62,7 +64,7 @@ def import_bigquery_service_account_impersonation():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"cast(cast( ('0x' || substr(TO_HEX(md5({s})), 18)) as int64) as numeric)"
+        return f"cast(cast( ('0x' || substr(TO_HEX(md5({s})), {CHECKSUM_HEXDIGITS})) as int64) as numeric) - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/clickhouse.py

@@ -6,6 +6,7 @@
     MD5_HEXDIGITS,
     CHECKSUM_HEXDIGITS,
     TIMESTAMP_PRECISION_POS,
+    CHECKSUM_OFFSET,
     BaseDialect,
     ThreadedDatabase,
     import_helper,
@@ -41,7 +42,7 @@ def import_clickhouse():
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
         substr_idx = 1 + MD5_HEXDIGITS - CHECKSUM_HEXDIGITS
-        return f"reinterpretAsUInt128(reverse(unhex(lowerUTF8(substr(hex(MD5({s})), {substr_idx})))))"
+        return f"reinterpretAsUInt128(reverse(unhex(lowerUTF8(substr(hex(MD5({s})), {substr_idx}))))) - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/databricks.py

@@ -21,6 +21,7 @@
 from data_diff.databases.base import (
     MD5_HEXDIGITS,
     CHECKSUM_HEXDIGITS,
+    CHECKSUM_OFFSET,
     BaseDialect,
     ThreadedDatabase,
     import_helper,
@@ -39,7 +40,7 @@ def import_databricks():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"cast(conv(substr(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16, 10) as decimal(38, 0))"
+        return f"cast(conv(substr(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16, 10) as decimal(38, 0)) - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/duckdb.py

@@ -29,6 +29,7 @@
     ConnectError,
     ThreadLocalInterpreter,
     TIMESTAMP_PRECISION_POS,
+    CHECKSUM_OFFSET
 )
 from data_diff.databases.base import MD5_HEXDIGITS, CHECKSUM_HEXDIGITS, Mixin_Schema
 from data_diff.queries.ast_classes import Func, Compilable, ITable
@@ -45,7 +46,7 @@ def import_duckdb():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"('0x' || SUBSTRING(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS},{CHECKSUM_HEXDIGITS}))::BIGINT"
+        return f"('0x' || SUBSTRING(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS},{CHECKSUM_HEXDIGITS}))::BIGINT - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/mssql.py

@@ -60,7 +60,8 @@ def normalize_number(self, value: str, coltype: FractionalType) -> str:
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"convert(bigint, convert(varbinary, '0x' + RIGHT(CONVERT(NVARCHAR(32), HashBytes('MD5', {s}), 2), {CHECKSUM_HEXDIGITS}), 1))"
+        offset = 549755813887
+        return f"convert(bigint, convert(varbinary, '0x' + RIGHT(CONVERT(NVARCHAR(32), HashBytes('MD5', {s}), 2), {CHECKSUM_HEXDIGITS}), 1)) - {offset}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/mysql.py

@@ -31,6 +31,7 @@
     MD5_HEXDIGITS,
     CHECKSUM_HEXDIGITS,
     TIMESTAMP_PRECISION_POS,
+    CHECKSUM_OFFSET,
     Mixin_Schema,
     Mixin_RandomSample,
 )
@@ -47,7 +48,7 @@ def import_mysql():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"cast(conv(substring(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16, 10) as unsigned)"
+        return f"cast(conv(substring(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16, 10) as unsigned) - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/oracle.py

@@ -27,6 +27,8 @@
     ConnectError,
     QueryError,
     Mixin_RandomSample,
+    CHECKSUM_OFFSET,
+    CHECKSUM_HEXDIGITS
 )
 from data_diff.databases.base import TIMESTAMP_PRECISION_POS
 
@@ -45,7 +47,7 @@ class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
         # standard_hash is faster than DBMS_CRYPTO.Hash
         # TODO: Find a way to use UTL_RAW.CAST_TO_BINARY_INTEGER ?
-        return f"to_number(substr(standard_hash({s}, 'MD5'), 18), 'xxxxxxxxxxxxxxx')"
+        return f"to_number(substr(standard_hash({s}, 'MD5'), {CHECKSUM_HEXDIGITS}), 'xxxxxxxxxxxxxxx') - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/postgresql.py

@@ -25,6 +25,7 @@
     CHECKSUM_HEXDIGITS,
     _CHECKSUM_BITSIZE,
     TIMESTAMP_PRECISION_POS,
+    CHECKSUM_OFFSET,
     Mixin_RandomSample,
 )
 
@@ -42,7 +43,7 @@ def import_postgresql():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"('x' || substring(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}))::bit({_CHECKSUM_BITSIZE})::bigint"
+        return f"('x' || substring(md5({s}), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}))::bit({_CHECKSUM_BITSIZE})::bigint - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)

data_diff/databases/presto.py

@@ -33,6 +33,7 @@
 from data_diff.databases.base import (
     MD5_HEXDIGITS,
     CHECKSUM_HEXDIGITS,
+    CHECKSUM_OFFSET,
     TIMESTAMP_PRECISION_POS,
 )
 
@@ -56,7 +57,7 @@ def import_presto():
 @attrs.define(frozen=False)
 class Mixin_MD5(AbstractMixin_MD5):
     def md5_as_int(self, s: str) -> str:
-        return f"cast(from_base(substr(to_hex(md5(to_utf8({s}))), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16) as decimal(38, 0))"
+        return f"cast(from_base(substr(to_hex(md5(to_utf8({s}))), {1+MD5_HEXDIGITS-CHECKSUM_HEXDIGITS}), 16) as decimal(38, 0)) - {CHECKSUM_OFFSET}"
 
 
 @attrs.define(frozen=False)