Hyukjin Kwon created SPARK-62072:
------------------------------------
Summary: An Observation's _name, _result and _error are its
attributes, its plan is built of its _name each time, its metrics are set at
the action, and a signal handler's Exception is caught by the plan compression
lookup, as PySpark Connect
Key: SPARK-62072
URL: https://issues.apache.org/jira/browse/SPARK-62072
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: 5.0.0
Reporter: Hyukjin Kwon
Follow-ups of SPARK-62009 (#853) and SPARK-62005 (#850). Differences of the
drop-in from pyspark-client 4.3.0:
1. The attributes of an Observation. PySpark Connect's Observation
(python/pyspark/sql/connect/observation.py:34-86) is a plain class whose
`__init__` sets `self._name`, `self._result` and `self._error`, which `get` and
`_on` read, and which its client sets at each action
(connect/client/core.py:1766-1799): the client reads
`observations[name]._result`, asserts it is not None, converts the metrics with
LiteralExpression._to_value and updates it in place, or gives the error of
collecting them to `_set_error`. The drop-in's Observation had them as
read-only getters of its native state:
- vars(Observation("o")) was {} where PySpark's is {'_name': 'o', '_result':
None, '_error': None};
- `o._name = ...`, `o._result = None` (which PySpark's `_on` then uses
again) and `o._error = exc` (which `get` then raises) raised AttributeError:
attribute '_name' of 'Observation' objects is not writable;
- each read of `_error` was a new exception, where PySpark's is the one
object that its client set;
- a reference cycle through the result dict (`o.get["self"] = o`) was never
freed, as the native state held the dict;
- the metrics were converted when `get` was read: a metric of a timestamp of
the year 1, whose conversion to a local datetime raises ValueError: year 0 is
out of range, made `get` raise where PySpark's action raises it and leaves
`get` empty ({}).
Once `_name` is an attribute that a program sets, the plan must read it as
PySpark's does (3).
2. A SIGINT handler that raises an Exception subclass while a call waits.
PySpark's lookup of the plan compression settings of the first plan
(_get_plan_compression_threshold_and_algorithm, core.py:2714-2740) catches
Exception, so the exception that a handler raises while the lookup waits is
caught, the compression is disabled, and the call goes on; its execution
interrupts the operation of a KeyboardInterrupt only (core.py:1943-1950), and
raises the exception of another handler at once. So the clean_cache of the exit
hook of a session that ran no plan, which looks the settings up first, needs
two signals of such a handler to end against a server that does not answer: in
a probe, PySpark's exit hook ended after 2 signals, the drop-in's after 1; and
the first action of such a session raised the second signal's exception in
PySpark, the first in the drop-in. The drop-in also interrupted the operation
of any handler's exception, so the next signal only ended the wait of the
interrupt: an action against a server that does not answer raised the second
signal's exception after 3 signals.
3. The name of an Observation at each build of its plan. PySpark's
CollectMetrics (python/pyspark/sql/connect/plan.py:1817-1837) sets
`str(self._observation._name)` in the name of its message each time it builds
the plan, after its input, and its `observations` name the Observation by that
str, by which the client sets the metrics; Observation._on does not read the
name. The drop-in's plan, which the Rust core builds, is of the name that the
Observation had at `observe`, so once a program can set `_name` (1):
- after `o._name = "other"`, the requests of its DataFrames would name the
first name, where PySpark's name "other", by which its metrics are set;
- a rename to another object, such as 5, None or b"n", which PySpark sends
as its str ("5", "None", "b'n'"), or to a name whose `__str__` raises or a str
that UTF-8 cannot encode, which fail each of PySpark's requests with that error
(the exception of `__str__`, or UnicodeEncodeError) until it is renamed again,
would change nothing;
- a name of another object set before `observe` would be taken as protobuf
takes a string field (bytes as their text, 5 as protobuf's TypeError of each
request), not as its str.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]