Exceptions that arise from a bulk DML call (including any recursive
DML operations in triggers that are fired as a direct result of the
call) are handled differently depending on where the original call
came from:
When errors occur because of a bulk DML call that originates directly
from the Apex DML statements, or if the all_or_none parameter of a database DML method was specified as true, the runtime
engine follows the “all or nothing” rule: during a single operation,
all records must be updated successfully or the entire operation rolls
back to the point immediately preceding the DML statement.
When errors occur because of a bulk DML call that originates from
the Web services API, the runtime engine attempts at least a partial save:
During the first attempt, the runtime engine processes all records.
Any record that generates an error due to issues such as validation
rules or unique index violations is set aside.
If there were errors during the first attempt, the runtime engine
makes a second attempt which includes only those records that did
not generate errors. All records that didn't generate an error during
the first attempt are processed, and if any record generates an error
(perhaps because of race conditions) it is also set aside.
If there were additional errors during the second attempt, the
runtime engine makes a third and final attempt which includes only
those records that did not generate errors during the first and second
attempts. If any record generates an error, the entire operation fails
with the error message, “Too many batch retries in the presence of Apex triggers and partial failures.”