diff --git a/Dockerfile b/Dockerfile
index 757d0c00..396ef25c 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -14,18 +14,18 @@ RUN pip install . -t ./python/lib/$runtime/site-packages
 # Remove *.pyc files
 RUN find ./python/lib/$runtime/site-packages -name \*.pyc -delete
 
-# Strip symbols from ddtrace's binaries.
-# TODO (AJ): remove when ddtrace fixes this upstream
-RUN find . -name '*.so' -exec strip -g {} \;
-
 # Remove botocore (40MB) to reduce package size. aws-xray-sdk
 # installs it, while it's already provided by the Lambda Runtime.
 RUN rm -rf ./python/lib/$runtime/site-packages/botocore*
 RUN rm -rf ./python/lib/$runtime/site-packages/setuptools
 RUN rm -rf ./python/lib/$runtime/site-packages/jsonschema/tests
 RUN find . -name 'libddwaf.so' -delete
+RUN rm -rf ./python/lib/$runtime/site-packages/urllib3
 RUN rm ./python/lib/$runtime/site-packages/ddtrace/appsec/_iast/_taint_tracking/*.so
 RUN rm ./python/lib/$runtime/site-packages/ddtrace/appsec/_iast/_stacktrace*.so
+RUN rm ./python/lib/$runtime/site-packages/ddtrace/internal/datadog/profiling/libdd_wrapper.so
+RUN rm ./python/lib/$runtime/site-packages/ddtrace/internal/datadog/profiling/ddup/_ddup.*.so
+RUN rm ./python/lib/$runtime/site-packages/ddtrace/internal/datadog/profiling/stack_v2/_stack_v2.*.so
 
 FROM scratch
 COPY --from=builder /build/python /
diff --git a/ci/input_files/build.yaml.tpl b/ci/input_files/build.yaml.tpl
index 68efeea8..449f6978 100644
--- a/ci/input_files/build.yaml.tpl
+++ b/ci/input_files/build.yaml.tpl
@@ -9,6 +9,7 @@ stages:
   - virtualenv venv
   - source venv/bin/activate
   - pip install .[dev]
+  - pip install poetry
 
 # This is for serverless framework
 .install-node: &install-node
@@ -53,7 +54,7 @@ check-layer-size ({{ $runtime.name }}-{{ $runtime.arch }}):
   dependencies:
     - build-layer ({{ $runtime.name }}-{{ $runtime.arch }})
   script: 
-    - PYTHON_VERSION={{ $runtime.python_version }} ./scripts/check_layer_size.sh
+    - PYTHON_VERSION={{ $runtime.python_version }} ARCH={{ $runtime.arch }} ./scripts/check_layer_size.sh
 
 lint python:
   stage: test
@@ -167,6 +168,7 @@ publish-pypi-package:
   stage: publish
   tags: ["arch:amd64"]
   image: registry.ddbuild.io/images/docker:20.10-py3
+  before_script: *python-before-script
   cache: []
   rules:
     - if: '$CI_COMMIT_TAG =~ /^v.*/'
diff --git a/ci/publish_pypi.sh b/ci/publish_pypi.sh
index d7ec78fd..c01df7ef 100755
--- a/ci/publish_pypi.sh
+++ b/ci/publish_pypi.sh
@@ -13,6 +13,14 @@ PYPI_TOKEN=$(aws ssm get-parameter \
     --out text)
 # Builds the lambda layer and upload to Pypi
 
+if [ -z "$CI_COMMIT_TAG" ]; then
+    printf "[Error] No CI_COMMIT_TAG found.\n"
+    printf "Exiting script...\n"
+    exit 1
+else
+    printf "Tag found in environment: $CI_COMMIT_TAG\n"
+fi
+
 # Clear previously built distributions
 if [ -d "dist" ]; then
     echo "Removing folder 'dist' to clear previously built distributions"
diff --git a/datadog_lambda/cold_start.py b/datadog_lambda/cold_start.py
index 9da02e78..9dcbec23 100644
--- a/datadog_lambda/cold_start.py
+++ b/datadog_lambda/cold_start.py
@@ -8,6 +8,7 @@
 _cold_start = True
 _proactive_initialization = False
 _lambda_container_initialized = False
+_tracer = None
 
 
 def set_cold_start(init_timestamp_ns):
@@ -18,6 +19,7 @@ def set_cold_start(init_timestamp_ns):
     global _cold_start
     global _lambda_container_initialized
     global _proactive_initialization
+    global _tracer
     if not _lambda_container_initialized:
         now = time.time_ns()
         if (now - init_timestamp_ns) // 1_000_000_000 > 10:
@@ -29,6 +31,7 @@ def set_cold_start(init_timestamp_ns):
         _cold_start = False
         _proactive_initialization = False
     _lambda_container_initialized = True
+    from ddtrace import tracer as _tracer
 
 
 def is_cold_start():
@@ -62,6 +65,9 @@ def __init__(self, module_name, full_file_path, start_time_ns, end_time_ns=None)
         self.start_time_ns = start_time_ns
         self.end_time_ns = end_time_ns
         self.children = []
+        self.context = None
+        if _lambda_container_initialized:
+            self.context = _tracer.context_provider.active()
 
 
 root_nodes: List[ImportNode] = []
@@ -70,10 +76,8 @@ def __init__(self, module_name, full_file_path, start_time_ns, end_time_ns=None)
 
 
 def reset_node_stacks():
-    global root_nodes
-    root_nodes = []
-    global import_stack
-    import_stack = []
+    root_nodes.clear()
+    import_stack.clear()
 
 
 def push_node(module_name, file_path):
@@ -183,7 +187,8 @@ def trace(self, root_nodes: List[ImportNode] = root_nodes):
         cold_start_span = self.create_cold_start_span(cold_start_span_start_time_ns)
         while root_nodes:
             root_node = root_nodes.pop()
-            self.trace_tree(root_node, cold_start_span)
+            parent = root_node.context or cold_start_span
+            self.trace_tree(root_node, parent)
         self.finish_span(cold_start_span, cold_start_span_end_time_ns)
 
     def trace_tree(self, import_node: ImportNode, parent_span):
diff --git a/pyproject.toml b/pyproject.toml
index 94b73d03..806485b6 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "datadog_lambda"
-version = "5.90.0"
+version = "5.91.0"
 description = "The Datadog AWS Lambda Library"
 authors = ["Datadog, Inc. <dev@datadoghq.com>"]
 license = "Apache-2.0"
diff --git a/scripts/check_layer_size.sh b/scripts/check_layer_size.sh
index f1c8c813..84752fa1 100755
--- a/scripts/check_layer_size.sh
+++ b/scripts/check_layer_size.sh
@@ -8,8 +8,9 @@
 # Compares layer size to threshold, and fails if below that threshold
 
 # 7 mb size limit
-MAX_LAYER_COMPRESSED_SIZE_KB=$(expr 7 \* 1024)
-MAX_LAYER_UNCOMPRESSED_SIZE_KB=$(expr 24 \* 1024)
+set -e
+MAX_LAYER_COMPRESSED_SIZE_KB=$(expr 4 \* 1024)
+MAX_LAYER_UNCOMPRESSED_SIZE_KB=$(expr 13 \* 1024)
 
 
 LAYER_FILES_PREFIX="datadog_lambda_py"
diff --git a/tests/test_cold_start.py b/tests/test_cold_start.py
index 65193e1d..2ce37e7c 100644
--- a/tests/test_cold_start.py
+++ b/tests/test_cold_start.py
@@ -1,10 +1,13 @@
+import os
 import time
 import unittest
-import datadog_lambda.cold_start as cold_start
+
 from sys import modules, meta_path
-import os
 from unittest.mock import MagicMock
 
+import datadog_lambda.cold_start as cold_start
+import datadog_lambda.wrapper as wrapper
+
 
 class TestColdStartTracingSetup(unittest.TestCase):
     def test_proactive_init(self):
@@ -234,3 +237,42 @@ def test_trace_ignore_libs(self):
         self.cold_start_tracer.trace(nodes)
         self.mock_activate.assert_called_once_with(self.mock_trace_ctx)
         self.assertEqual(self.output_spans, ["node_0", "unittest_cold_start"])
+
+
+def test_lazy_loaded_package_imports(monkeypatch):
+
+    spans = []
+
+    def finish(span):
+        spans.append(span)
+
+    monkeypatch.setattr(wrapper.tracer, "_on_span_finish", finish)
+    monkeypatch.setattr(wrapper, "is_new_sandbox", lambda: True)
+    monkeypatch.setattr("datadog_lambda.wrapper.dd_tracing_enabled", True)
+    monkeypatch.setenv(
+        "DD_COLD_START_TRACE_SKIP_LIB", "ddtrace.contrib.logging,datadog_lambda.wrapper"
+    )
+    monkeypatch.setenv("DD_MIN_COLD_START_DURATION", "0")
+
+    @wrapper.datadog_lambda_wrapper
+    def handler(event, context):
+        import tabnanny
+
+    lambda_context = MagicMock()
+    lambda_context.invoked_function_arn = (
+        "arn:aws:lambda:us-west-1:123457598159:function:python-layer-test:1"
+    )
+
+    handler.cold_start_tracing = True
+    handler({}, lambda_context)
+
+    function_span = import_span = None
+    for span in spans:
+        if span.resource == "tabnanny":
+            import_span = span
+        elif span.name == "aws.lambda":
+            function_span = span
+
+    assert function_span is not None
+    assert import_span is not None
+    assert import_span.parent_id == function_span.span_id