Improve config options for missing and dropping db columns.

Implement actual drop functionality.

Signed-off-by: Jason Volk <jason@zemos.net>

Cargo.lock

@@ -5079,6 +5079,7 @@ dependencies = [
  "criterion",
  "ctor",
  "futures",
+ "itertools 0.14.0",
  "log",
  "minicbor",
  "minicbor-serde",

src/core/config/mod.rs

@@ -1288,19 +1288,20 @@ pub struct Config {
 	#[serde(default = "default_rocksdb_stats_level")]
 	pub rocksdb_stats_level: u8,
 
-	/// Erases data no longer reachable in the current schema. The developers
-	/// expect this to be set to true which simplifies the schema and prevents
-	/// accumulation of old schemas remaining in the codebase forever. If this
-	/// is set to false, old columns which are not described in the current
-	/// schema will be ignored rather than erased, leaking their space.
+	/// Ignores the list of dropped columns set by developers.
 	///
-	/// This can be set to false when moving between versions in ways which are
-	/// not recommended or otherwise forbidden, or for diagnostic and
-	/// development purposes; requiring preservation across such movements.
+	/// This should be set to true when knowingly moving between versions in
+	/// ways which are not recommended or otherwise forbidden, or for
+	/// diagnostic and development purposes; requiring preservation across such
+	/// movements.
 	///
-	/// default: true
-	#[serde(default = "true_fn")]
-	pub rocksdb_drop_missing_columns: bool,
+	/// The developer's list of dropped columns is meant to safely reduce space
+	/// by erasing data no longer in use. If this is set to true that storage
+	/// will not be reclaimed as intended.
+	///
+	/// default: false
+	#[serde(default)]
+	pub rocksdb_never_drop_columns: bool,
 
 	/// This is a password that can be configured that will let you login to the
 	/// server bot account (currently `@conduit`) for emergency troubleshooting

src/database/Cargo.toml

@@ -57,6 +57,7 @@ async-channel.workspace = true
 const-str.workspace = true
 ctor.workspace = true
 futures.workspace = true
+itertools.workspace = true
 log.workspace = true
 minicbor.workspace = true
 minicbor-serde.workspace = true

src/database/engine/open.rs

@@ -4,15 +4,12 @@ use std::{
 	sync::{Arc, atomic::AtomicU32},
 };
 
+use itertools::Itertools;
 use rocksdb::{ColumnFamilyDescriptor, Options};
-use tuwunel_core::{Result, debug, implement, info, warn};
+use tuwunel_core::{Result, debug, debug_warn, implement, info, warn};
 
 use super::{
-	Db, Engine,
-	cf_opts::cf_options,
-	context,
-	db_opts::db_options,
-	descriptor::{self, Descriptor},
+	Db, Engine, cf_opts::cf_options, context, db_opts::db_options, descriptor::Descriptor,
 	repair::repair,
 };
 use crate::{Context, or_else};
@@ -31,7 +28,7 @@ pub(crate) async fn open(ctx: Arc<Context>, desc: &[Descriptor]) -> Result<Arc<S
 		&ctx.row_cache.lock().expect("row cache locked"),
 	)?;
 
-	let cfds = Self::configure_cfds(&ctx, &db_opts, desc)?;
+	let (cfds, dropped) = Self::configure_cfds(&ctx, &db_opts, desc)?;
 	let num_cfds = cfds.len();
 	debug!("Configured {num_cfds} column descriptors...");
 
@@ -50,6 +47,13 @@ pub(crate) async fn open(ctx: Arc<Context>, desc: &[Descriptor]) -> Result<Arc<S
 	}
 	.or_else(or_else)?;
 
+	if !config.rocksdb_read_only && !config.rocksdb_secondary {
+		for name in &dropped {
+			debug!("Deleting dropped column {name:?} ...");
+			db.drop_cf(name).or_else(or_else)?;
+		}
+	}
+
 	info!(
 		columns = num_cfds,
 		sequence = %db.latest_sequence_number(),
@@ -74,59 +78,86 @@ fn configure_cfds(
 	ctx: &Arc<Context>,
 	db_opts: &Options,
 	desc: &[Descriptor],
-) -> Result<Vec<ColumnFamilyDescriptor>> {
+) -> Result<(Vec<ColumnFamilyDescriptor>, Vec<String>)> {
 	let server = &ctx.server;
 	let config = &server.config;
 	let path = &config.database_path;
 	let existing = Self::discover_cfs(path, db_opts);
 
-	let creating = desc
-		.iter()
-		.filter(|desc| !existing.contains(desc.name));
-
+	// Found columns which are not described.
 	let missing = existing
 		.iter()
 		.filter(|&name| name != "default")
 		.filter(|&name| !desc.iter().any(|desc| desc.name == name));
 
+	// Described columns which are not found.
+	let creating = desc
+		.iter()
+		.filter(|desc| !desc.dropped)
+		.filter(|desc| !existing.contains(desc.name));
+
+	// Found columns which are described as dropped.
+	let dropping = desc
+		.iter()
+		.filter(|desc| desc.dropped)
+		.filter(|desc| existing.contains(desc.name))
+		.filter(|_| !config.rocksdb_never_drop_columns);
+
+	// Described dropped columns which are no longer found.
+	let dropped = desc
+		.iter()
+		.filter(|desc| desc.dropped)
+		.filter(|desc| !existing.contains(desc.name));
+
 	debug!(
 		existing = existing.len(),
 		described = desc.len(),
 		missing = missing.clone().count(),
+		dropped = dropped.clone().count(),
 		creating = creating.clone().count(),
+		dropping = dropping.clone().count(),
 		"Discovered database columns"
 	);
 
 	missing.clone().for_each(|name| {
-		debug!("Found unrecognized column {name:?} in existing database.");
+		debug_warn!("Found undescribed column {name:?} in existing database.");
+	});
+
+	dropped.map(|desc| desc.name).for_each(|name| {
+		debug!("Previously dropped column {name:?} no longer found in database.");
 	});
 
 	creating.map(|desc| desc.name).for_each(|name| {
 		debug!("Creating new column {name:?} not previously found in existing database.");
 	});
 
-	let missing_descriptors = missing
+	dropping
 		.clone()
-		.filter(|_| config.rocksdb_drop_missing_columns)
-		.map(|_| descriptor::DROPPED);
+		.map(|desc| desc.name)
+		.for_each(|name| {
+			warn!(
+				"Column {name:?} has been scheduled for deletion. Storage may not appear \
+				 reclaimed until further restart or compaction."
+			);
+		});
 
-	let cfopts: Vec<_> = desc
-		.iter()
-		.copied()
-		.chain(missing_descriptors)
-		.map(|ref desc| cf_options(ctx, db_opts.clone(), desc))
-		.collect::<Result<_>>()?;
+	let dropping_names: Vec<_> = dropping
+		.clone()
+		.map(|desc| desc.name)
+		.map(ToOwned::to_owned)
+		.collect();
 
 	let cfds: Vec<_> = desc
 		.iter()
-		.map(|desc| desc.name)
-		.map(ToOwned::to_owned)
-		.chain(missing.cloned())
-		.zip(cfopts.into_iter())
-		.map(|(name, opts)| ColumnFamilyDescriptor::new(name, opts))
-		.collect();
+		.filter(|desc| !desc.dropped)
+		.chain(dropping)
+		.copied()
+		.inspect(|desc| debug!(name = desc.name, "Described column"))
+		.map(|desc| Ok((desc.name.to_owned(), cf_options(ctx, db_opts.clone(), &desc)?)))
+		.map_ok(|(name, opts)| ColumnFamilyDescriptor::new(name, opts))
+		.collect::<Result<_>>()?;
 
-	Ok(cfds)
+	Ok((cfds, dropping_names))
 }
 
 #[implement(Engine)]

src/database/maps.rs

@@ -16,6 +16,7 @@ pub(super) fn open(engine: &Arc<Engine>) -> Result<Maps> { open_list(engine, MAP
 #[tracing::instrument(name = "maps", level = "debug", skip_all)]
 pub(super) fn open_list(engine: &Arc<Engine>, maps: &[Descriptor]) -> Result<Maps> {
 	maps.iter()
+		.filter(|desc| !desc.dropped)
 		.map(|desc| Ok((desc.name, Map::open(engine, desc.name)?)))
 		.collect()
 }
@@ -165,6 +166,10 @@ pub(super) static MAPS: &[Descriptor] = &[
 		name: "roomid_joinedcount",
 		..descriptor::RANDOM_SMALL
 	},
+	Descriptor {
+		name: "roomid_maxremotepowerlevel",
+		..descriptor::RANDOM_SMALL
+	},
 	Descriptor {
 		name: "roomid_pduleaves",
 		..descriptor::RANDOM_SMALL
@@ -447,8 +452,4 @@ pub(super) static MAPS: &[Descriptor] = &[
 		name: "userroomid_notificationcount",
 		..descriptor::RANDOM
 	},
-	Descriptor {
-		name: "roomid_maxremotepowerlevel",
-		..descriptor::RANDOM_SMALL
-	},
 ];

src/database/mod.rs

@@ -1,4 +1,4 @@
-#![type_length_limit = "8192"]
+#![type_length_limit = "65536"]
 
 extern crate rust_rocksdb as rocksdb;
 

tuwunel-example.toml

@@ -1091,17 +1091,18 @@
 #
 #rocksdb_stats_level = 1
 
-# Erases data no longer reachable in the current schema. The developers
-# expect this to be set to true which simplifies the schema and prevents
-# accumulation of old schemas remaining in the codebase forever. If this
-# is set to false, old columns which are not described in the current
-# schema will be ignored rather than erased, leaking their space.
+# Ignores the list of dropped columns set by developers.
 #
-# This can be set to false when moving between versions in ways which are
-# not recommended or otherwise forbidden, or for diagnostic and
-# development purposes; requiring preservation across such movements.
+# This should be set to true when knowingly moving between versions in
+# ways which are not recommended or otherwise forbidden, or for
+# diagnostic and development purposes; requiring preservation across such
+# movements.
 #
-#rocksdb_drop_missing_columns = true
+# The developer's list of dropped columns is meant to safely reduce space
+# by erasing data no longer in use. If this is set to true that storage
+# will not be reclaimed as intended.
+#
+#rocksdb_never_drop_columns = false
 
 # This is a password that can be configured that will let you login to the
 # server bot account (currently `@conduit`) for emergency troubleshooting