diff --git a/compiler/compile-config/compile-config-2.02-learning_inference b/compiler/compile-config/compile-config-2.02-learning_inference
index 7702c4968..74fdfda3e 100755
--- a/compiler/compile-config/compile-config-2.02-learning_inference
+++ b/compiler/compile-config/compile-config-2.02-learning_inference
@@ -58,6 +58,11 @@ if .deepdive_.execution.processes | has("process/grounding/combine_factorgraph")
                     $DEEPDIVE_SAMPLER_ARGS \\
                     #
 
+            \(if $deepdive.sampler.partitions > 1 then "
+            # make sure no stale inference result is laying around unless inference is done right after learning above
+            rm -f \"$results_dir\"/inference_result.out.text
+            " else "" end)
+
             # TODO maybe this database interaction for the weights after
             # learning each shard is unnecessary if we have the sampler checkpoint
             # them directly, and we could also keep max two copies
@@ -191,10 +196,7 @@ if .deepdive_.execution.processes | has("process/grounding/combine_factorgraph")
             for pid in $DEEPDIVE_FACTORGRAPH_SHARDS; do
                 cd \"$DEEPDIVE_RESULTS_DIR\"/$pid
                 cat inference_result.out.text |
-                # restoring shard ID to the vids
-                awk -v SHARD_BASE=$(($pid << 48)) '
-                    {printf \"%s\\t%s\\t%s\\n\", SHARD_BASE + $1, $2, $3}
-                ' |
+                pid=$pid deepdive env restore_partitioned_vids |
                 DEEPDIVE_LOAD_FORMAT=tsv \\
                 deepdive load \(deepdiveInferenceResultVariablesTable | @sh) /dev/stdin
             done
diff --git a/database/db-driver/greenplum/db-unload b/database/db-driver/greenplum/db-unload
index ef1f02f9c..ee7883563 100755
--- a/database/db-driver/greenplum/db-unload
+++ b/database/db-driver/greenplum/db-unload
@@ -6,15 +6,16 @@
 # parallel connections, which MAY RESULT IN INCORRECT RESULT UNLESS QUERY IS
 # TOTALLY ORDERED BY COLUMNS WITH NO DUPLICATE VALUES!
 #
-# GPFDIST_PORT_BASE (defaults to 10000) is the base port number for the gpfdist
-# processes.
+# GPFDIST_PORT_BASE is the base port number for the gpfdist processes.
+# (defaults to a random port in range 10000-59900)
 # GPFDIST_MAX_LENGTH (defaults to 1MiB) may need to be increased for larger
 # records to be unloaded through gpfdist.
 # GPFDIST_LOG_PATH can be set to a path prefix for logging.
 ##
 set -eu
 
-: ${GPFDIST_DISABLE:=} ${GPFDIST_PORT_BASE:=10000} ${GPFDIST_MAX_LENGTH:=$((2**20))} ${GPFDIST_LOG_PATH:=}
+: ${GPFDIST_DISABLE:=} ${GPFDIST_MAX_LENGTH:=$((2**20))} ${GPFDIST_LOG_PATH:=}
+: ${GPFDIST_PORT_BASE:=$(( 100*($RANDOM % 500 + 100) ))} # randomized default base ports
 
 [[ $# -gt 0 ]] || usage "$0" "Missing QUERY"
 query=$1; shift
@@ -70,12 +71,14 @@ case $format in
                 sinkDir=$(dirname "$sink") sinkName=$(basename "$sink")
                 mkdir -p "$sinkDir"
                 spawn_gpfdist() {
+                    # skip over ports already taken
+                    while nc -z "$sinkHost" "$port"; do let ++port; done
                     gpfdist -p $port \
                         ${GPFDIST_MAX_LENGTH:+-m $GPFDIST_MAX_LENGTH} \
                         ${GPFDIST_LOG_PATH:+-l ${GPFDIST_LOG_PATH}.$i.log} &>/dev/null \
                         "$@" &
                     echo $! >>"$tmpDir"/gpfdist-$i.pid
-                    # TODO cope with gpfdist failure, e.g., port number collision
+                    # TODO cope with gpfdist failure, e.g., when another process suddenly tried to listen to the picked port number
                 }
                 if ${DEEPDIVE_SHOW_PROGRESS:-true}; then
                     spawn_gpfdist -d "$tmpDir"
diff --git a/inference/restore_partitioned_vids b/inference/restore_partitioned_vids
new file mode 100755
index 000000000..519a7f63f
--- /dev/null
+++ b/inference/restore_partitioned_vids
@@ -0,0 +1,7 @@
+#!/usr/bin/perl -w
+use strict;
+my $SHARD_BASE = $ENV{pid} << 48;
+while (<>) {
+    my ($vid, @rest) = split(" ");
+    print join("\t", ($SHARD_BASE | $vid), @rest), "\n";
+}
diff --git a/runner/compute-driver/local/compute-execute b/runner/compute-driver/local/compute-execute
index c72e0445f..c23f1b86f 100755
--- a/runner/compute-driver/local/compute-execute
+++ b/runner/compute-driver/local/compute-execute
@@ -131,7 +131,8 @@ all_finishes_ok() {
         }
     done
 }
+# the order of below is important. if we don't check unload first and unload fails, we are stuck
+[[ ${#pids_unload[@]}  -eq 0 ]] || all_finishes_ok "deepdive-unload"                 "${pids_unload[@]}"
 [[ ${#pids_command[@]} -eq 0 ]] || all_finishes_ok "command=$(escape4sh "$command")" "${pids_command[@]}"
 [[ ${#pids_load[@]}    -eq 0 ]] || all_finishes_ok "deepdive-load"                   "${pids_load[@]}"
-[[ ${#pids_unload[@]}  -eq 0 ]] || all_finishes_ok "deepdive-unload"                 "${pids_unload[@]}"
 wait  # until everything is done ##############################################
diff --git a/stage.sh b/stage.sh
index 137dd16a4..be3accf22 100755
--- a/stage.sh
+++ b/stage.sh
@@ -142,6 +142,7 @@ generate-wrapper-for-libdirs          "$STAGE_DIR"/util/sampler-$cmd \
 done
 stage inference/deepdive-model                                    util/
 stage inference/run-sampler                                       util/
+stage inference/restore_partitioned_vids                          util/
 
 # Stanford CoreNLP utilities
 stage util/nlp/deepdive-corenlp                                   util/