Darren Weber created ARROW-5324:
-----------------------------------
Summary: plasma API requests
Key: ARROW-5324
URL: https://issues.apache.org/jira/browse/ARROW-5324
Project: Apache Arrow
Issue Type: Improvement
Reporter: Darren Weber
Copied from [https://github.com/apache/arrow/issues/4318]
While working with the plasma API to create/seal an object for a table, using a
custom object-ID, it would help to have a convenience API to get the size of
the table.
The following code might help to illustrate the request and notes below:
```python
if not parquet_path:
parquet_path = f"./data/dataset_
{size}.parquet"
if not plasma_path:
plasma_path = f"./data/dataset_\{size}
.plasma"
try:
plasma_client = plasma.connect(plasma_path)
except:
plasma_client = None
if plasma_client:
table_id = plasma.ObjectID(bytes(parquet_path[:20], encoding='utf8'))
try:
table = plasma_client.get(table_id, timeout_ms=4000)
if table.__name__ == 'ObjectNotAvailable':
raise ValueError('Failed to get plasma object')
except ValueError:
table = pq.read_table(parquet_path, use_threads=True)
plasma_client.create_and_seal(table_id, table)
```
The use case is a workflow something like this:
- process-A
- generate a pandas DataFrame `df`
- save the `df` to parquet, using pyarrow.parquet, with a unique parquet path
- (this process will not save directly to plasma)
- process-B
- get the data from plasma or load it into plasma from the parquet file
- use the unique parquet path to generate a unique object-ID
Notes:
- `plasma_client.put` for the same data-table is not idempotent, it generates
unique object-ID values that are not based on any hash of the data payload, so
every put saves a new object-ID; could it use a data hash for idempotent puts?
e.g.
```python
In : plasma_client.put(table)
ObjectID(666625fcb60959d23b6bfc739f88816da29e04d6)
In : plasma_client.put(table)
ObjectID(d2a4662999db30177b090f9fc2bf6b28687d2f8d)
In : plasma_client.put(table)
ObjectID(b2928ad786de2fdb74d374055597f6e7bd97fd61)
In : hash(table)
TypeError: unhashable type: 'pyarrow.lib.Table'
```
- In process-B, when the data is not already in plasma, it reads data from a
parquet file into a pyarrow.Table and then needs an object-ID and the table
size to use plasma `client.create_and_seal` but it's not easy to get the table
size - this might be related to github issue #2707 - it might be ideal if the
`client.create_and_seal` accepts responsibility for the size of the object to
be created when given a pyarrow data object like a table.
- when the plasma store does not have the object, it could have a default
timeout rather than hang indefinitely, and it's a bit clumsy to return an
object that is not easily checked with `isinstance` and it could be better to
have an exception handling pattern (or something like the requests 404 patterns
and options?)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
