PL/SQL :: Sample Function Returns Many Times Of Expected Value

The SAMPLE clause in the select statement works well in most cases, but we found in some instances the result is way off - between 200% to 700% discrepancy has been observed.

For example, we have thee tables with the following results:

Table1: 495,365,317 rows (20 cols, unique primary key present), SAMPLE ( 0.002018712182064212 ) returns 41,499 (about four times off - we expected about 10,000)
Table2: 3,350,864,539 rows ( 5 cols, unique primary key present), SAMPLE ( 0.00029843044634040336 ) returns 9,835 (this is good as it is close to 10,000)
Table3: 6,974,724,543 rows ( 5 cols, no unique primary key present), SAMPLE ( 0.00014337483779250091 ) returns 58,789 (about six times off - we expected about 10,000)

The tables got billions of rows, and that is why we want to do sampling. The sample percentage rate is computed to return about 10,000 rows in all three tables.On Table3, we ran the sampling three times in one occasion, and we got "58,570", "24,575" and "24,561"

I expected +/- 20% of variance, but 200% to 700% seems to be way too much.Once again, I stress that it does work well in most cases (another 3.4 billion table and numerous smaller tables we tested were well within +/- 5 percent of the target).I noted the presence of a primary key above because I read an article saying that the SAMPLE function relies on the existence of a primary key (which does not quite explain the examples above).Is this kind of spread something we should expect or is it a bug? Is the sampling rate too small for such large tables?