Add config variable for timeline offloading (#9421)

Adds a configuration variable for timeline offloading support. The added
pageserver-global config option controls whether the pageserver
automatically offloads timelines during compaction.

Therefore, already offloaded timelines are not affected by this, nor is
the manual testing endpoint.

This allows the rollout of timeline offloading to be driven by the
storage team.

Part of #8088

libs/pageserver_api/src/config.rs

@@ -102,6 +102,7 @@ pub struct ConfigToml {
     pub ingest_batch_size: u64,
     pub max_vectored_read_bytes: MaxVectoredReadBytes,
     pub image_compression: ImageCompressionAlgorithm,
+    pub timeline_offloading: bool,
     pub ephemeral_bytes_per_memory_kb: usize,
     pub l0_flush: Option<crate::models::L0FlushConfig>,
     pub virtual_file_io_mode: Option<crate::models::virtual_file::IoMode>,
@@ -385,6 +386,7 @@ impl Default for ConfigToml {
                 NonZeroUsize::new(DEFAULT_MAX_VECTORED_READ_BYTES).unwrap(),
             )),
             image_compression: (DEFAULT_IMAGE_COMPRESSION),
+            timeline_offloading: false,
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
             l0_flush: None,
             virtual_file_io_mode: None,

pageserver/src/config.rs

@@ -164,6 +164,9 @@ pub struct PageServerConf {
 
     pub image_compression: ImageCompressionAlgorithm,
 
+    /// Whether to offload archived timelines automatically
+    pub timeline_offloading: bool,
+
     /// How many bytes of ephemeral layer content will we allow per kilobyte of RAM.  When this
     /// is exceeded, we start proactively closing ephemeral layers to limit the total amount
     /// of ephemeral data.
@@ -321,6 +324,7 @@ impl PageServerConf {
             ingest_batch_size,
             max_vectored_read_bytes,
             image_compression,
+            timeline_offloading,
             ephemeral_bytes_per_memory_kb,
             l0_flush,
             virtual_file_io_mode,
@@ -364,6 +368,7 @@ impl PageServerConf {
             ingest_batch_size,
             max_vectored_read_bytes,
             image_compression,
+            timeline_offloading,
             ephemeral_bytes_per_memory_kb,
 
             // ------------------------------------------------------------

pageserver/src/tenant.rs

@@ -2187,7 +2187,8 @@ impl Tenant {
                             .iter()
                             .any(|(_id, tl)| tl.get_ancestor_timeline_id() == Some(*timeline_id))
                     };
-                    let can_offload = can_offload && has_no_unoffloaded_children;
+                    let can_offload =
+                        can_offload && has_no_unoffloaded_children && self.conf.timeline_offloading;
                     if (is_active, can_offload) == (false, false) {
                         None
                     } else {

pageserver/src/tenant/timeline.rs

@@ -1565,6 +1565,7 @@ impl Timeline {
     }
 
     /// Checks if the internal state of the timeline is consistent with it being able to be offloaded.
+    ///
     /// This is neccessary but not sufficient for offloading of the timeline as it might have
     /// child timelines that are not offloaded yet.
     pub(crate) fn can_offload(&self) -> bool {

test_runner/regress/test_timeline_archive.py

@@ -119,6 +119,10 @@ def test_timeline_archive(neon_env_builder: NeonEnvBuilder, shard_count: int):
 
 @pytest.mark.parametrize("manual_offload", [False, True])
 def test_timeline_offloading(neon_env_builder: NeonEnvBuilder, manual_offload: bool):
+    if not manual_offload:
+        # (automatic) timeline offloading defaults to false for now
+        neon_env_builder.pageserver_config_override = "timeline_offloading = true"
+
     env = neon_env_builder.init_start()
     ps_http = env.pageserver.http_client()