diff --git a/app/plugin/CustomSectionAPI-ti.ts b/app/plugin/CustomSectionAPI-ti.ts
index 37f361f3..6d7b8853 100644
--- a/app/plugin/CustomSectionAPI-ti.ts
+++ b/app/plugin/CustomSectionAPI-ti.ts
@@ -7,6 +7,7 @@ import * as t from "ts-interface-checker";
 export const ColumnToMap = t.iface([], {
   "name": "string",
   "title": t.opt(t.union("string", "null")),
+  "description": t.opt(t.union("string", "null")),
   "type": t.opt("string"),
   "optional": t.opt("boolean"),
   "allowMultiple": t.opt("boolean"),
diff --git a/app/server/lib/SafePythonComponent.ts b/app/server/lib/SafePythonComponent.ts
index 9d0eb87a..1518d8cc 100644
--- a/app/server/lib/SafePythonComponent.ts
+++ b/app/server/lib/SafePythonComponent.ts
@@ -41,6 +41,7 @@ export class SafePythonComponent extends BaseComponent {
       importMount: this._tmpDir,
       logTimes: true,
       logMeta: this._logMeta,
+      preferredPythonVersion: '3',
     });
   }
 
diff --git a/sandbox/grist/imports/fixtures/strange_dates.xlsx b/sandbox/grist/imports/fixtures/strange_dates.xlsx
index 8219e1fe..4efdc848 100644
Binary files a/sandbox/grist/imports/fixtures/strange_dates.xlsx and b/sandbox/grist/imports/fixtures/strange_dates.xlsx differ
diff --git a/sandbox/grist/imports/test_import_csv.py b/sandbox/grist/imports/import_csv_test.py
similarity index 100%
rename from sandbox/grist/imports/test_import_csv.py
rename to sandbox/grist/imports/import_csv_test.py
diff --git a/sandbox/grist/imports/test_import_json.py b/sandbox/grist/imports/import_json_test.py
similarity index 100%
rename from sandbox/grist/imports/test_import_json.py
rename to sandbox/grist/imports/import_json_test.py
diff --git a/sandbox/grist/imports/import_utils.py b/sandbox/grist/imports/import_utils.py
index 07d83b77..37671568 100644
--- a/sandbox/grist/imports/import_utils.py
+++ b/sandbox/grist/imports/import_utils.py
@@ -1,17 +1,17 @@
 """
 Helper functions for import plugins
 """
-import sys
 import itertools
 import logging
 import os
 
-# Include /thirdparty into module search paths, in particular for messytables.
-sys.path.append('/thirdparty')
-
 import six
 from six.moves import zip
 
+if six.PY2:
+  raise RuntimeError("Imports should use a Python 3 environment")
+
+
 log = logging.getLogger(__name__)
 
 # Get path to an imported file.
diff --git a/sandbox/grist/imports/import_xls.py b/sandbox/grist/imports/import_xls.py
index ebd15150..2f3adfe3 100644
--- a/sandbox/grist/imports/import_xls.py
+++ b/sandbox/grist/imports/import_xls.py
@@ -5,8 +5,10 @@ and returns a object formatted so that it can be used by grist for a bulk add re
 import logging
 
 import messytables
-import openpyxl
 import six
+import openpyxl
+from openpyxl.utils.datetime import from_excel
+from openpyxl.worksheet import _reader
 from six.moves import zip
 
 import parse_data
@@ -15,6 +17,19 @@ from imports import import_utils
 log = logging.getLogger(__name__)
 
 
+# Some strange Excel files have values that are marked as dates but are invalid as dates.
+# Normally, openpyxl converts these to `#VALUE!`, but keeping the original value is better.
+# In the case where this was encountered, the original value is a large number:
+# the 6281228502068 in test_excel_strange_dates.
+# Here we monkeypatch openpyxl to keep the original value.
+def new_from_excel(value, *args, **kwargs):
+  try:
+    return from_excel(value, *args, **kwargs)
+  except (ValueError, OverflowError):
+    return value
+_reader.from_excel = new_from_excel
+
+
 def import_file(file_source):
   path = import_utils.get_path(file_source["path"])
   parse_options, tables = parse_file(path)
diff --git a/sandbox/grist/imports/test_import_xls.py b/sandbox/grist/imports/import_xls_test.py
similarity index 98%
rename from sandbox/grist/imports/test_import_xls.py
rename to sandbox/grist/imports/import_xls_test.py
index 86d5da1d..a905ef9a 100644
--- a/sandbox/grist/imports/test_import_xls.py
+++ b/sandbox/grist/imports/import_xls_test.py
@@ -133,7 +133,8 @@ class TestImportXLS(unittest.TestCase):
         {'id': 'f', 'type': 'Date'},
         {'id': 'g', 'type': 'Any'},
         {'id': 'h', 'type': 'Date'},
-        {'id': 'i', 'type': 'Date'},
+        {'id': 'i', 'type': 'Any'},
+        {'id': 'j', 'type': 'Numeric'},
       ],
       'table_data': [
         [u'21:14:00'],
@@ -144,7 +145,8 @@ class TestImportXLS(unittest.TestCase):
         [-2207347200.0],
         [u'7/4/1776'],
         [205286400.0],
-        [-2209161600.0],
+        ['00:00:00'],
+        [6281228502068],
       ],
     }])
 
diff --git a/sandbox/grist/imports/test_messytables.py b/sandbox/grist/imports/messytables_test.py
similarity index 100%
rename from sandbox/grist/imports/test_messytables.py
rename to sandbox/grist/imports/messytables_test.py
diff --git a/sandbox/grist/imports/test_imports.py b/sandbox/grist/imports/test_imports.py
new file mode 100644
index 00000000..25893b6c
--- /dev/null
+++ b/sandbox/grist/imports/test_imports.py
@@ -0,0 +1,41 @@
+"""
+Imports only run in Python 3 sandboxes, so the tests here only run in Python 3.
+The test files in this directory have been renamed to look like 'import_xls_test.py' instead
+of 'test_import_xls.py' so that they're not discovered automatically by default.
+`load_tests` below then discovers that pattern directly, but only in Python 3.
+This allows the tests to be skipped without having to specify a pattern when discovering tests.
+The downside is that if you *do* want to specify a pattern, that probably won't work.
+
+The reason for this method is that there's a bug in Python 2's unittest module
+regarding packages and directories that we must handle.
+This means that in Python 2, I wasn't able to prevent files like 'test_import_xls.py'
+from being discovered automatically regardless of what I did in `load_tests`.
+
+Compare https://docs.python.org/3.9/library/unittest.html#load-tests-protocol
+     vs https://docs.python.org/2/library/unittest.html#load-tests-protocol
+
+from "If discovery is started" on both pages. Note in particular:
+
+  > Changed in version 3.5: Discovery no longer checks package names for matching pattern
+    due to the impossibility of package names matching the default pattern.
+
+The reason for skipping entire files from being discovered instead of skipping TestCase classes
+is that just importing the test file will fail with an error, both because we manually raise
+an exception and because dependencies are missing.
+"""
+
+import os
+
+import six
+
+
+def load_tests(loader, standard_tests, _pattern):
+  if six.PY2:
+    return standard_tests
+
+  this_dir = os.path.join(os.path.dirname(__file__))
+  package_tests = list(loader.discover(start_dir=this_dir, pattern='*_test.py'))
+  if len(package_tests) < 3:
+    raise Exception("Expected more import tests to be discovered")
+  standard_tests.addTests(package_tests)
+  return standard_tests
diff --git a/sandbox/requirements.txt b/sandbox/requirements.txt
index 49a32d01..d956f6c9 100644
--- a/sandbox/requirements.txt
+++ b/sandbox/requirements.txt
@@ -3,22 +3,12 @@ asttokens==2.0.5
 backports.functools-lru-cache==1.6.4
 chardet==4.0.0
 enum34==1.1.10
-et-xmlfile==1.0.1
-html5lib==1.1
 iso8601==0.1.12
-jdcal==1.4.1
-json_table_schema==0.2.1
 lazy_object_proxy==1.6.0
-lxml==4.6.3                # used in csv plugin only?
-messytables==0.15.2
-openpyxl==2.6.4
 python_dateutil==2.8.2
-python_magic==0.4.12
 roman==2.0.0
 singledispatch==3.6.2
 six==1.16.0
 sortedcontainers==2.4.0
-webencodings==0.5
 wrapt==1.12.1
-xlrd==1.2.0
 unittest-xml-reporting==2.0.0
diff --git a/sandbox/requirements3.txt b/sandbox/requirements3.txt
index f0e39250..7f016bfc 100644
--- a/sandbox/requirements3.txt
+++ b/sandbox/requirements3.txt
@@ -20,8 +20,8 @@ json_table_schema==0.2.1
 lazy_object_proxy==1.6.0
 lxml==4.6.3                # used in csv plugin only?
 messytables==0.15.2
-openpyxl==2.6.4
 python_dateutil==2.8.2
+openpyxl==3.0.10
 python_magic==0.4.12
 roman==2.0.0
 singledispatch==3.6.2