diff --git a/stats-backfill/src/backfill_repos.rs b/stats-backfill/src/backfill_repos.rs
index 59afd9f..1694609 100644
--- a/stats-backfill/src/backfill_repos.rs
+++ b/stats-backfill/src/backfill_repos.rs
@@ -178,6 +178,8 @@ pub struct Args {
     pub max_rss_mb: u64,
     /// Directory under which each tier-3 disk spill creates a fresh subdir.
     pub spill_dir: PathBuf,
+    /// Hosts to skip entirely (never picked as candidates).
+    pub skip_hosts: Vec<pds::Hostname>,
 }
 
 pub async fn run(db: &Db, args: Args, ua: &str) -> Result<(), Error> {
@@ -201,6 +203,16 @@ pub async fn run(db: &Db, args: Args, ua: &str) -> Result<(), Error> {
     let spill_dir = Arc::new(args.spill_dir);
     let progress = Arc::new(Progress::default());
 
+    let skip_hosts: std::collections::HashSet<pds::Hostname> =
+        args.skip_hosts.into_iter().collect();
+    if !skip_hosts.is_empty() {
+        tracing::info!(
+            count = skip_hosts.len(),
+            hosts = ?skip_hosts.iter().collect::<Vec<_>>(),
+            "skipping hosts (--skip)"
+        );
+    }
+
     let ticker = spawn_progress_ticker(progress.clone(), args.max_rss_mb);
 
     let mut active: JoinSet<Result<(), Error>> = JoinSet::new();
@@ -211,7 +223,7 @@ pub async fn run(db: &Db, args: Args, ua: &str) -> Result<(), Error> {
         // kick off as many per-host workers as allowed
         while active.len() < args.host_workers_limit {
             let candidates = db.eligible_hosts().await?;
-            let Some(host) = pick_next_host(&candidates, &active_hosts) else {
+            let Some(host) = pick_next_host(&candidates, &active_hosts, &skip_hosts) else {
                 break;
             };
 
@@ -354,10 +366,11 @@ fn ensure_nofile_for_big_repos(big_repos_limit: usize) -> Result<(), Error> {
 fn pick_next_host(
     candidates: &[HostCandidate],
     active: &std::collections::HashSet<pds::Hostname>,
+    skip: &std::collections::HashSet<pds::Hostname>,
 ) -> Option<pds::Hostname> {
     candidates
         .iter()
-        .filter(|c| !active.contains(&c.host))
+        .filter(|c| !active.contains(&c.host) && !skip.contains(&c.host))
         .max_by_key(|c| c.listhosts_account_count.unwrap_or(0))
         .map(|c| c.host.clone())
 }
diff --git a/stats-backfill/src/main.rs b/stats-backfill/src/main.rs
index e8e2d83..eb591df 100644
--- a/stats-backfill/src/main.rs
+++ b/stats-backfill/src/main.rs
@@ -114,6 +114,13 @@ struct BackfillReposArgs {
     /// it's often tmpfs (RAM-backed), which defeats the purpose.
     #[arg(long, default_value = "./local/spill")]
     spill_dir: PathBuf,
+
+    /// skip these PDS hosts entirely — they won't be picked as candidates.
+    /// Repeatable: `--skip foo.example --skip bar.example`. Useful for hosts
+    /// that are known-broken or known-unreachable in ways we can't auto-
+    /// detect (e.g. operator has disabled getRepo network-wide).
+    #[arg(long)]
+    skip: Vec<pds::Hostname>,
 }
 
 #[tokio::main]
@@ -176,5 +183,6 @@ fn scan_args(args: BackfillReposArgs) -> backfill_repos::Args {
         big_repos_limit: args.big_repos_limit,
         max_rss_mb: args.max_rss_mb,
         spill_dir: args.spill_dir,
+        skip_hosts: args.skip,
     }
 }