From ee8901c1c201031e49d08b3fbd03db57daeadeda Mon Sep 17 00:00:00 2001
From: Zongheng Yang <zongheng.y@gmail.com>
Date: Fri, 28 Oct 2022 12:47:53 -0700
Subject: [PATCH 1/2] Add roadmap.

---
 ROADMAP.md | 52 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)
 create mode 100644 ROADMAP.md

diff --git a/ROADMAP.md b/ROADMAP.md
new file mode 100644
index 00000000000..e2de58cbe48
--- /dev/null
+++ b/ROADMAP.md
@@ -0,0 +1,52 @@
+# SkyPilot Roadmap
+
+## v0.3
+
+### Managed Spot
+- Minimize the cost of the controller
+  - Support spot controller on existing cluster (or local cluster)
+  - Reducing the fixed cost of the controller (e.g., allow setting controller VM type)
+  - Increasing parallelism (number of concurrent jobs)
+- Pushing the scale (e.g., support a high number of pending/concurrent jobs)
+- Framework-specific guides to add checkpointing/reloading using SkyPilot Storage
+
+### Smarter Optimizer
+- Fine-grained optimizer: pick by cheapest zone order
+- Better consider data egress time/cost
+  - Consider buckets/Storage objects in file_mounts
+- Optimizing the data placement for SkyPilot Storage local uploads
+  - Use the optimizer to decide the bucket location
+
+### Programmatic API
+- Refactor/extend the current API to *make it easy to programmatically use SkyPilot*
+- Expose core classes in docs
+
+### More clouds
+- Refactoring of interfaces to ease adding new clouds
+- IBM Cloud
+
+### On-prem
+- Design for switching between cloud and on-prem
+- Explore/design of "local mode" to run SkyPilot tasks locally
+
+### Faster launching speed
+- Consider a more minimal image
+- Azure speed investigation
+
+### k8s support
+- Ray-on-k8s backend
+  - To figure out: Launch a new k8s cluster? Launch SkyPilot Tasks to an existing k8s cluster?
+
+### Cost: Optimization, Tracking, and Reporting
+- Track and show costs related to a job/cluster
+- For managed spot jobs, track and show %savings vs. on-demand
+- Optimizer: take into account disk costs
+
+### Serverless
+- Design and prototype of a "serverless jobs" submission API and CLI
+  - Initial use case: hundreds of hyperparameter tuning trials
+
+### Backend
+- Support heterogeneous node types in a cluster (e.g., in RL, CPU actor(s) and GPU learner(s) in the same cluster)
+- Support CPUs as resource requirements
+- General robustness/UX improvements

From 49f14ea6c4f56884b8ccc2b67faa15adbb90f030 Mon Sep 17 00:00:00 2001
From: Zongheng Yang <zongheng.y@gmail.com>
Date: Sun, 30 Oct 2022 20:31:17 -0700
Subject: [PATCH 2/2] Update ROADMAP.md

---
 ROADMAP.md | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/ROADMAP.md b/ROADMAP.md
index e2de58cbe48..bd62d08f87c 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -1,13 +1,18 @@
 # SkyPilot Roadmap
 
+This doc lists general directions of interest to facilitate community contributions. 
+
+Note that
+- This list is not meant to be comprehensive (i.e., new work items of interest may pop up)
+- Even though listed under a specific version, not all items need to be completed before we ship that version (i.e., some items can go into future versions)
+
 ## v0.3
 
 ### Managed Spot
 - Minimize the cost of the controller
-  - Support spot controller on existing cluster (or local cluster)
+  - Support running spot controller on an existing/local cluster
   - Reducing the fixed cost of the controller (e.g., allow setting controller VM type)
-  - Increasing parallelism (number of concurrent jobs)
-- Pushing the scale (e.g., support a high number of pending/concurrent jobs)
+- Supporting a higher number of pending/concurrent jobs
 - Framework-specific guides to add checkpointing/reloading using SkyPilot Storage
 
 ### Smarter Optimizer
@@ -21,11 +26,13 @@
 - Refactor/extend the current API to *make it easy to programmatically use SkyPilot*
 - Expose core classes in docs
 
-### More clouds
+### Support more clouds
 - Refactoring of interfaces to ease adding new clouds
 - IBM Cloud
+- Explore support for low-cost clouds (e.g., lambda labs/runpod/jarvis labs)
 
 ### On-prem
+- Robustify the on-prem feature
 - Design for switching between cloud and on-prem
 - Explore/design of "local mode" to run SkyPilot tasks locally