Fix precision loss for large CQL timestamp values

[Copilot]

Pre-review checklist

• I have split my patch into logically separate commits.
• All commit messages clearly explain what they change and why.
• I added relevant tests for new features and bug fixes.
• All commits compile, pass static checks and pass test.
• PR description sums up the changes and reasons why they should be introduced.
• I have provided docstrings for the public items that I want to introduce.
• I have adjusted the documentation in ./docs/source/.

Description

CQL timestamp values (signed milliseconds from epoch) lose precision when they are further than ~300 years from Unix epoch due to floating-point conversions during serialization/deserialization. This causes issues where a timestamp retrieved from the database cannot be used to query for the same row.

Problem

The driver was converting int64 milliseconds to floating-point values during serialization/deserialization, causing precision loss for timestamps more than ~300 years from the Unix epoch (the maximum number of milliseconds that can be exactly represented in a double).

Example scenario:

• Insert timestamp "2300-01-01 00:00:00.001"
• Retrieve it back: "2300-01-01 00:00:00.000999" (precision lost)
• Use retrieved value in a query: fails because it's serialized as "2300-01-01 00:00:00.000"

Root Causes

1. Deserialization divides int64 milliseconds by 1000.0 (float), losing precision
2. Serialization uses floating-point arithmetic (*1e3, /1e3) to calculate milliseconds
3. Both Python and Cython implementations had this issue

Solution

Implemented integer-only arithmetic throughout the timestamp conversion path:

1. Added new conversion functions:

   • datetime_from_timestamp_ms() in cassandra/util.py (Python path)
   • datetime_from_timestamp_ms() in cassandra/cython_utils.pyx (Cython path)
   • Both functions use integer arithmetic exclusively to avoid precision loss

2. Updated DateType serialization/deserialization:

   • cassandra/cqltypes.py: Changed DateType.deserialize() to call the new function directly on milliseconds (no division)
   • cassandra/cqltypes.py: Changed DateType.serialize() to use integer operations (*1000, //1000) instead of float operations (*1e3, /1e3)
   • cassandra/deserializers.pyx: Updated Cython deserializer to use the new integer-based function

3. Code cleanup:

   • Removed unnecessary negative remainder check from datetime_from_timestamp_ms() - Python's modulo operator always returns non-negative results for positive divisors

4. Added comprehensive test coverage:

   • Created tests/unit/test_timestamp_precision.py with 4 tests:
     • Large timestamp roundtrip (year 2300)
     • Various large timestamps including negative values (year 1640 BCE)
     • Small timestamps to ensure no regression
     • The exact scenario from the issue report

Testing

• ✅ All new tests pass (4/4)
• ✅ All existing DateType tests pass
• ✅ All unit type tests pass (62 passed, 1 skipped)
• ✅ No regressions detected
• ✅ Issue scenario validated - timestamps like "2300-01-01 00:00:00.001" now maintain full precision through serialization/deserialization cycles
• ✅ Negative timestamp handling verified to work correctly

The changes are minimal and surgical, affecting only the timestamp conversion logic without altering any other functionality.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Large values of CQL type timestamp lose precision due to some floating-point conversions along the way</issue_title>
<issue_description>timestamp CQL values are integers (signed milliseconds from epoch), but they get converted to double somewhere along the way, so they lose precision if they are further than ~300 years from the Unix epoch. (~300 years is the max number of milliseconds which can be exactly represented in a double).

In particular, you can e.g. select a timestamp key from a table, then use the obtained key to query the table for that particular key, and the query will return an empty result.

Consider:

    keyspace_name = "ks"
    table_name = "ks"
    session.execute(f"DROP KEYSPACE IF EXISTS {keyspace_name}")
    session.execute(f"""
        CREATE KEYSPACE IF NOT EXISTS {keyspace_name}
        WITH replication = {{'class': 'SimpleStrategy', 'replication_factor': 1}}
    """)
    session.execute(f"""
        CREATE TABLE {keyspace_name}.{table_name} (
            pk timestamp PRIMARY KEY
        )
    """)
    test_timestamp = "2300-01-01 00:00:00.001"
    insert = session.execute(f"INSERT INTO {keyspace_name}.{table_name}(pk) VALUES('{test_timestamp}')")

    scan_result = session.execute(f"SELECT * FROM {keyspace_name}.{table_name}")
    scan_rows = list(scan_result)
    assert len(scan_rows) == 1

    retrieved_timestamp = scan_rows[0].pk
    logger.debug(f"Original timestamp: {test_timestamp}")
    logger.debug(f"Retrieved timestamp: {retrieved_timestamp}")

    key_rows = list(session.execute(f"SELECT * FROM {keyspace_name}.{table_name} WHERE pk = '{test_timestamp}'"))
    assert len(key_rows) == 1

    key_select = session.prepare(f"SELECT * FROM {keyspace_name}.{table_name} WHERE pk = ?")
    key_rows = list(session.execute(key_select, [retrieved_timestamp]))
    assert len(key_rows) == 1 # FAIL

    session.execute(f"DROP KEYSPACE {keyspace_name}")

This will give the log output:

2025-09-11 23:12:40,117 - DEBUG - Original timestamp: 2300-01-01 00:00:00.001
2025-09-11 23:12:40,117 - DEBUG - Retrieved timestamp: 2300-01-01 00:00:00.000999

And the test will fail on the assert marked with the FAIL comment, because the actual key is 2300-01-01 00:00:00.001, the key converted to Python native types is 2300-01-01 00:00:00.001 - epsilon, and the key sent back with the last SELECT is 2300-01-01 00:00:00.000.</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes Large values of CQL type timestamp lose precision due to some floating-point conversions along the way #532

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[mykaul]

@copilot please complete the work on the PR

[Copilot]

@copilot please complete the work on the PR

The PR is now complete. All changes have been implemented and tested:

Changes made (commit fbab5cd):

• Added integer-based timestamp conversion functions to avoid floating-point precision loss
• Updated serialization/deserialization in both Python and Cython code paths
• Added comprehensive test suite that validates the fix

Testing results:

• New timestamp precision tests: ✅ All 4 tests pass
• Existing DateType tests: ✅ Pass
• All unit type tests (63 tests): ✅ 62 passed, 1 skipped
• Issue scenario validation: ✅ Timestamps far from epoch now maintain full precision

The fix ensures that timestamps like "2300-01-01 00:00:00.001" can be correctly round-tripped through serialization/deserialization without losing precision.

[Lorak-mmk]

@copilot When can remainder be negative? Even for negative values it is positive:

>>> -2500 % 1000
500

[Copilot]

You're absolutely right! Python's modulo operator always returns a non-negative result when the divisor is positive. The negative check was unnecessary. Removed in commit 93ad0f2.