fix(unparser): Fix BigQuery timestamp literal format in SQL unparsing

[sgrebnov]

Which issue does this PR close?

The default Dialect::timestamp_with_tz_to_string uses dt.to_string() which produces timestamps with a space before the TimeZone offset. This causes filter pushdown to fail when unparsing timestamp predicates for BigQuery.

2016-08-06 20:05:00 +00:00 <- invalid for BigQuery:
invalid timestamp: '2016-08-06 20:05:00 +00:00'; while executing the filter on column 'startTime' (query) (sqlstate: [0, 0, 0, 0, 0], vendor_code: -2147483648)

BigQuery rejects this format. Per the BigQuery timestamp docs, the offset must be attached directly to the time:

2016-08-06 20:05:00+00:00 <- valid

What changes are included in this PR?

Following similar to DuckDB pattern/fix override timestamp_with_tz_to_string for BigQueryDialect to produce valid timestamp format

Before (default dt.to_string()):

CAST('2016-08-06 20:05:00 +00:00' AS TIMESTAMP)  -- BigQuery error

After (%:z format):

CAST('2016-08-06 20:05:00+00:00' AS TIMESTAMP)  -- valid BigQuery timestamp

Are these changes tested?

Added unit test and manual e2e test with Google BigQuery instance.

Are there any user-facing changes?

No

[nuno-faria]

LGTM, thanks @sgrebnov.

I wonder if timestamp_with_tz_to_string should use the version without the space as the default, since I think according to ISO 8601 it shouldn't be there.

I tested with PostgreSQL/SQLite/MySQL and they all support the version without the space as well.

[sgrebnov]

dt.to_string()

@nuno-faria - that is a great point, I think we should be using dt.to_rfc3339() by default, which should resolve DuckDB and BigQuery issues as well so we even don't need custom overrides for them. WDYT?

/// Returns an RFC 3339 and ISO 8601 date and time string such as `1996-12-19T16:39:57-08:00`.
    #[cfg(feature = "alloc")]
    #[must_use]
    pub fn to_rfc3339(&self) -> String {
        // For some reason a string with a capacity less than 32 is ca 20% slower when benchmarking.
        let mut result = String::with_capacity(32);
        let naive = self.overflowing_naive_local();
        let offset = self.offset.fix();
        write_rfc3339(&mut result, naive, offset, SecondsFormat::AutoSi, false)
            .expect("writing rfc3339 datetime to string should never fail");
        result
    }

[nuno-faria]

dt.to_string()

@nuno-faria - that is a great point, I think we should be using dt.to_rfc3339() by default, which should resolve DuckDB and BigQuery issues as well so we even don't need custom overrides for them. WDYT?

/// Returns an RFC 3339 and ISO 8601 date and time string such as `1996-12-19T16:39:57-08:00`.
    #[cfg(feature = "alloc")]
    #[must_use]
    pub fn to_rfc3339(&self) -> String {
        // For some reason a string with a capacity less than 32 is ca 20% slower when benchmarking.
        let mut result = String::with_capacity(32);
        let naive = self.overflowing_naive_local();
        let offset = self.offset.fix();
        write_rfc3339(&mut result, naive, offset, SecondsFormat::AutoSi, false)
            .expect("writing rfc3339 datetime to string should never fail");
        result
    }

Sounds good to me. Should we do on this PR or leave as a follow up?

[sgrebnov]

@nuno-faria - I propose to merge this one as-is and I'll create another one to switch to to_rfc3339 next. This is to ensure we have a backup plan to revert added switch to to_rfc3339 if needed (unlikely)

[AndreaBozzo]

Looks nice

[alamb]

Thanks @nuno-faria and @sgrebnov and @AndreaBozzo