diff --git a/.github/workflows/pull_request.yaml b/.github/workflows/pull_request.yaml
index 4636b0edee..e729437185 100644
--- a/.github/workflows/pull_request.yaml
+++ b/.github/workflows/pull_request.yaml
@@ -37,7 +37,7 @@ jobs:
     needs: lint
     runs-on: self-hosted
     container:
-      image: apachepegasus/ci-env
+      image: registry.cn-beijing.aliyuncs.com/apachepegasus/thirdparties-bin:ubuntu1804
       env:
         CCACHE_DIR: /tmp/ccache/pegasus
         CCACHE_MAXSIZE: 10G
@@ -45,14 +45,19 @@ jobs:
         # Place ccache compilation intermediate results in host memory, that's shared among containers.
         - /tmp/ccache/pegasus:/tmp/ccache/pegasus
       # Read docs at https://docs.docker.com/storage/tmpfs/ for more details of using tmpfs in docker.
-      options: --mount type=tmpfs,destination=/tmp/pegasus,tmpfs-size=10737418240 --cap-add=SYS_PTRACE
+      options: --mount type=tmpfs,destination=/tmp/pegasus --cap-add=SYS_PTRACE
+    defaults:
+      run:
+        shell: bash
+        working-directory: /root/rdsn
     steps:
-      - uses: actions/checkout@v2
-        with:
-          fetch-depth: 1
+      - name: Clone rdsn source
+        working-directory: /root
+        run: |
+          git clone --depth=1 https://hub.fastgit.org/XiaoMi/rdsn.git
       - name: Unpack prebuilt third-parties
         if: contains(github.event.pull_request.labels.*.name, 'thirdparty-modified') == false
-        run: unzip /root/pegasus-thirdparty-output.zip -d ./thirdparty
+        run: unzip /root/thirdparties-bin.zip -d ./thirdparty
       - name: Rebuild third-parties
         if: contains(github.event.pull_request.labels.*.name, 'thirdparty-modified')
         working-directory: thirdparty
@@ -63,4 +68,6 @@ jobs:
       - name: Compilation
         run: ./run.sh build -c --skip_thirdparty
       - name: Unit Testing
-        run: ./run.sh test --skip_thirdparty
+        run: |
+          export LD_LIBRARY_PATH=/root/rdsn/thirdparty/output/lib:/usr/lib/jvm/java-8-openjdk-amd64/jre/lib/amd64/server          
+          ./run.sh test --skip_thirdparty
diff --git a/src/replica/replica.h b/src/replica/replica.h
index 5886f7ee6b..f79d14fcdd 100644
--- a/src/replica/replica.h
+++ b/src/replica/replica.h
@@ -287,6 +287,15 @@ class replica : public serverlet<replica>, public ref_counter, public replica_ba
     void notify_learn_completion();
     error_code apply_learned_state_from_private_log(learn_state &state);
 
+    // Prepares in-memory mutations for the replica's learning.
+    // Returns false if there's no delta data in cache (aka prepare-list).
+    bool prepare_cached_learn_state(const learn_request &request,
+                                    decree learn_start_decree,
+                                    decree local_committed_decree,
+                                    /*out*/ remote_learner_state &learner_state,
+                                    /*out*/ learn_response &response,
+                                    /*out*/ bool &delayed_replay_prepare_list);
+
     // Gets the position where this round of the learning process should begin.
     // This method is called on primary-side.
     // TODO(wutao1): mark it const
diff --git a/src/replica/replica_learn.cpp b/src/replica/replica_learn.cpp
index ca8c79f544..23dceb27d5 100644
--- a/src/replica/replica_learn.cpp
+++ b/src/replica/replica_learn.cpp
@@ -437,62 +437,14 @@ void replica::on_learn(dsn::message_ex *msg, const learn_request &request)
     response.last_committed_decree = local_committed_decree;
     response.err = ERR_OK;
 
-    // set prepare_start_decree when to-be-learn state is covered by prepare list,
-    // note min_decree can be NOT present in prepare list when list.count == 0
-    if (learn_start_decree > _prepare_list->min_decree() ||
-        (learn_start_decree == _prepare_list->min_decree() && _prepare_list->count() > 0)) {
-        if (learner_state.prepare_start_decree == invalid_decree) {
-            // start from (last_committed_decree + 1)
-            learner_state.prepare_start_decree = local_committed_decree + 1;
-
-            cleanup_preparing_mutations(false);
-
-            // the replayed prepare msg needs to be AFTER the learning response msg
-            // to reduce probability that preparing messages arrive remote early than
-            // learning response msg.
-            delayed_replay_prepare_list = true;
-
-            ddebug("%s: on_learn[%016" PRIx64
-                   "]: learner = %s, set prepare_start_decree = %" PRId64,
-                   name(),
-                   request.signature,
-                   request.learner.to_string(),
-                   local_committed_decree + 1);
-        }
-
-        response.prepare_start_decree = learner_state.prepare_start_decree;
-    } else {
-        learner_state.prepare_start_decree = invalid_decree;
-    }
-
-    // only learn mutation cache in range of [learn_start_decree, prepare_start_decree),
-    // in this case, the state on the PS should be contiguous (+ to-be-sent prepare list)
-    if (response.prepare_start_decree != invalid_decree) {
-        binary_writer writer;
-        int count = 0;
-        for (decree d = learn_start_decree; d < response.prepare_start_decree; d++) {
-            auto mu = _prepare_list->get_mutation_by_decree(d);
-            dassert(mu != nullptr, "mutation must not be nullptr, decree = %" PRId64 "", d);
-            mu->write_to(writer, nullptr);
-            count++;
-        }
-        response.type = learn_type::LT_CACHE;
-        response.state.meta = writer.get_buffer();
-        ddebug("%s: on_learn[%016" PRIx64 "]: learner = %s, learn mutation cache succeed, "
-               "learn_start_decree = %" PRId64 ", prepare_start_decree = %" PRId64 ", "
-               "learn_mutation_count = %d, learn_data_size = %d",
-               name(),
-               request.signature,
-               request.learner.to_string(),
-               learn_start_decree,
-               response.prepare_start_decree,
-               count,
-               response.state.meta.length());
-    }
-
     // learn delta state or checkpoint
-    // in this case, the state on the PS is still incomplete
-    else {
+    bool should_learn_cache = prepare_cached_learn_state(request,
+                                                         learn_start_decree,
+                                                         local_committed_decree,
+                                                         learner_state,
+                                                         response,
+                                                         delayed_replay_prepare_list);
+    if (!should_learn_cache) {
         if (learn_start_decree > _app->last_durable_decree()) {
             ddebug("%s: on_learn[%016" PRIx64 "]: learner = %s, choose to learn private logs, "
                    "because learn_start_decree(%" PRId64 ") > _app->last_durable_decree(%" PRId64
@@ -1004,6 +956,69 @@ void replica::on_learn_reply(error_code err, learn_request &&req, learn_response
     }
 }
 
+bool replica::prepare_cached_learn_state(const learn_request &request,
+                                         decree learn_start_decree,
+                                         decree local_committed_decree,
+                                         /*out*/ remote_learner_state &learner_state,
+                                         /*out*/ learn_response &response,
+                                         /*out*/ bool &delayed_replay_prepare_list)
+{
+    // set prepare_start_decree when to-be-learn state is covered by prepare list,
+    // note min_decree can be NOT present in prepare list when list.count == 0
+    if (learn_start_decree > _prepare_list->min_decree() ||
+        (learn_start_decree == _prepare_list->min_decree() && _prepare_list->count() > 0)) {
+        if (learner_state.prepare_start_decree == invalid_decree) {
+            // start from (last_committed_decree + 1)
+            learner_state.prepare_start_decree = local_committed_decree + 1;
+
+            cleanup_preparing_mutations(false);
+
+            // the replayed prepare msg needs to be AFTER the learning response msg
+            // to reduce probability that preparing messages arrive remote early than
+            // learning response msg.
+            delayed_replay_prepare_list = true;
+
+            ddebug("%s: on_learn[%016" PRIx64
+                   "]: learner = %s, set prepare_start_decree = %" PRId64,
+                   name(),
+                   request.signature,
+                   request.learner.to_string(),
+                   local_committed_decree + 1);
+        }
+
+        response.prepare_start_decree = learner_state.prepare_start_decree;
+    } else {
+        learner_state.prepare_start_decree = invalid_decree;
+    }
+
+    // only learn mutation cache in range of [learn_start_decree, prepare_start_decree),
+    // in this case, the state on the PS should be contiguous (+ to-be-sent prepare list)
+    if (response.prepare_start_decree != invalid_decree) {
+        binary_writer writer;
+        int count = 0;
+        for (decree d = learn_start_decree; d < response.prepare_start_decree; d++) {
+            auto mu = _prepare_list->get_mutation_by_decree(d);
+            dassert(mu != nullptr, "mutation must not be nullptr, decree = %" PRId64 "", d);
+            mu->write_to(writer, nullptr);
+            count++;
+        }
+        response.type = learn_type::LT_CACHE;
+        response.state.meta = writer.get_buffer();
+        ddebug("%s: on_learn[%016" PRIx64 "]: learner = %s, learn mutation cache succeed, "
+               "learn_start_decree = %" PRId64 ", prepare_start_decree = %" PRId64 ", "
+               "learn_mutation_count = %d, learn_data_size = %d",
+               name(),
+               request.signature,
+               request.learner.to_string(),
+               learn_start_decree,
+               response.prepare_start_decree,
+               count,
+               response.state.meta.length());
+        return true;
+    }
+    return false;
+}
+
 void replica::on_copy_remote_state_completed(error_code err,
                                              size_t size,
                                              uint64_t copy_start_time,