Resubmission: Save output dataframe as HDF5, not Pickle

canine/orchestrator.py

@@ -8,12 +8,12 @@
 from .adapters import AbstractAdapter, ManualAdapter, FirecloudAdapter
 from .backends import AbstractSlurmBackend, LocalSlurmBackend, RemoteSlurmBackend, DummySlurmBackend, TransientGCPSlurmBackend, TransientImageSlurmBackend
 from .localization import AbstractLocalizer, BatchedLocalizer, LocalLocalizer, RemoteLocalizer, NFSLocalizer
-from .utils import check_call
+from .utils import check_call, pandas_read_hdf5_buffered, pandas_write_hdf5_buffered
 import yaml
 import numpy as np
 import pandas as pd
 from agutil import status_bar
-version = '0.8.1'
+version = '0.9.0'
 
 ADAPTERS = {
     'Manual': ManualAdapter,
@@ -416,6 +416,8 @@ def make_output_DF(self, batch_id, outputs, cpu_time, prev_acct, localizer = Non
                     # that don't receive any transformation with transformed columns
                     df["outputs"] = df["outputs"].agg({ **self.output_map, **identity_map })
             except:
+                df = pd.DataFrame()
+                print("Error generating output dataframe; see stack trace for details.", file = sys.stderr)
                 traceback.print_exc()
 
 
@@ -426,11 +428,11 @@ def make_output_DF(self, batch_id, outputs, cpu_time, prev_acct, localizer = Non
         # save DF to disk
         if isinstance(localizer, AbstractLocalizer):
             with localizer.transport_context() as transport:
-                dest = localizer.reserve_path("results.k9df.pickle").remotepath
+                dest = localizer.reserve_path("results.k9df.hdf5").remotepath
                 if not transport.isdir(os.path.dirname(dest)):
                     transport.makedirs(os.path.dirname(dest))
                 with transport.open(dest, 'wb') as w:
-                    df.to_pickle(w, compression=None)
+                    pandas_write_hdf5_buffered(df, buf = w, key = "results")
         return df
 
     def submit_batch_job(self, entrypoint_path, compute_env, extra_sbatch_args = {}) -> int:
@@ -459,7 +461,7 @@ def job_avoid(self, localizer: AbstractLocalizer, overwrite: bool = False) -> in
         Succeeded jobs are skipped. Failed jobs are reset and rerun
         """
         with localizer.transport_context() as transport:
-            df_path = localizer.reserve_path("results.k9df.pickle").remotepath
+            df_path = localizer.reserve_path("results.k9df.hdf5").remotepath
 
             #remove all output if specified
             if overwrite:
@@ -472,8 +474,8 @@ def job_avoid(self, localizer: AbstractLocalizer, overwrite: bool = False) -> in
             if transport.exists(df_path):
                 try:
                     # load in results and job spec dataframes
-                    with transport.open(df_path) as r:
-                        r_df = pd.read_pickle(r, compression=None)
+                    with transport.open(df_path, mode = "rb") as r:
+                        r_df = pandas_read_hdf5_buffered(key = "results", buf = r)
                     js_df = pd.DataFrame.from_dict(self.job_spec, orient = "index").rename_axis(index = "_job_id")
 
                     if r_df.empty or \
@@ -537,8 +539,8 @@ def job_avoid(self, localizer: AbstractLocalizer, overwrite: bool = False) -> in
 
                     return np.count_nonzero(~fail_idx)
                 except (ValueError, OSError) as e:
-                    print(e)
-                    print("Overwriting output and aborting job avoidance.")
+                    print("Cannot recover preexisting task outputs: " + str(e), file = sys.stderr)
+                    print("Overwriting output and aborting job avoidance.", file = sys.stderr)
                     transport.rmtree(localizer.staging_dir)
                     transport.makedirs(localizer.staging_dir)
                     return 0

canine/test/test_orchestrator.py

@@ -246,7 +246,7 @@ def test_cmd(self):
                 shell=True
             )
 
-    @with_timeout(180)
+    @with_timeout(300)
     def test_xargs(self):
         """
         Runs a quick test of the xargs orchestrator

canine/utils.py

@@ -12,6 +12,7 @@
 import paramiko
 import shutil
 import time
+import pandas as pd
 import requests
 
 def isatty(*streams: typing.IO) -> bool:
@@ -315,3 +316,29 @@ def gcp_hourly_cost(mtype: str, preemptible: bool = False, ssd_size: int = 0, hd
     )
 
 # rmtree_retry removed in favor of AbstractTransport.rmtree
+
+def pandas_write_hdf5_buffered(df: pd.DataFrame, key: str, buf: io.BufferedWriter):
+    """
+	Write a Pandas dataframe in HDF5 format to a buffer.
+    """
+    with pd.HDFStore(
+      "/dev/null",
+      mode = "w",
+      driver = "H5FD_CORE",
+      driver_core_backing_store = 0
+    ) as store:
+        store["results"] = df
+        buf.write(store._handle.get_file_image())
+
+def pandas_read_hdf5_buffered(key: str, buf: io.BufferedReader) -> pd.DataFrame:
+    """
+	Read a Pandas dataframe in HDF5 format from a buffer.
+    """
+    with pd.HDFStore(
+      "dummy_hdf5",
+      mode = "r",
+      driver = "H5FD_CORE",
+      driver_core_backing_store = 0,
+      driver_core_image = buf.read()
+    ) as store:
+        return store[key]

setup.py

@@ -56,7 +56,8 @@
         'google-api-python-client>=1.7.11',
         'docker>=4.1.0',
         'psutil>=5.6.7',
-        'port-for>=0.4'
+        'port-for>=0.4',
+        'tables>=3.6.1'
     ],
     classifiers = [
         "Development Status :: 4 - Beta",