Record ID written to state file as string

[untsakas]

I have two tables selected from Oracle database. For one of these ID column value for last record is written back to state file as text

---
last_records:
  app.application_log_vw:
    id: '163431515.0'

Next run fails for this reason with

OCIError: ORA-01722: invalid number

Only difference between the two tables is in ID having little different datatype declaration. One having NUMBER(38) and the problematic one only NUMBER.

Ruby gem versions

ruby-oci8-2.2.11
activerecord-oracle_enhanced-adapter-6.1.6
activerecord-6.1.4.6
activesupport-6.1.4.6
activerecord-oracle_enhanced-adapter-6.1.6
fluent-plugin-sql-2.3.0

[untsakas]

I can now confirm, that adding size to datatype declaration fixes the issue.

[Watson1978]

This breaks specifically for update_column/primary-key columns declared as bare NUMBER (no precision/scale).
The oracle_enhanced adapter casts an unconstrained NUMBER to a non-integer numeric type, whereas NUMBER(38) (scale 0) is cast to Integer.
That difference is what makes only one of the two tables fail.

From there it surfaces in the plugin's state handling.
In TableElement#read_attribute_for_serialization, values that aren't msgpack-serializable and aren't Time fall through to v.to_s.
An Integer is kept as-is, but the non-integer numeric is stringified to "163431515.0" and persisted to the state file as a quoted string.
On the next run, emit_next_records feeds that value straight back into the bind:

relation.where("#{@update_column} > ?", last_update_value)

So Oracle ends up comparing the NUMBER column against the string literal '163431515.0', which it tries to coerce and rejects with ORA-01722: invalid number.

Workaround (confirmed by reporter): declare a precision/scale on the column (e.g. NUMBER(38)) so the adapter returns an Integer. Selecting an integer-typed update_column should also avoid it.

Possible fix: the plugin could normalize the stored last value back to a numeric type before re-binding, rather than persisting/reusing the to_s form. Keeping this open as a bug to track that.