Wouldn't this work:

diff --git a/src/libstore/sqlite.cc b/src/libstore/sqlite.cc
index 06abfb90b..efe15b21b 100644
--- a/src/libstore/sqlite.cc
+++ b/src/libstore/sqlite.cc
@@ -249,7 +249,7 @@ void handleSQLiteBusy(const SQLiteBusy & e, time_t & nextWarning)
     if (now > nextWarning) {
         nextWarning = now + 10;
         logWarning({
-            .msg = HintFmt(e.what())
+            .msg = e.info().msg
         });
     }

Yeah please try @edolstra's fix. We want to treat the underlying disease (double rendering of error message) not the symptom (double level prefix).

(off topic)

underlying disease

Also what's causing starvation for over 10 seconds? Maybe registerValidPaths shouldn't hold a transaction while parsing derivations? Or maybe the backoffs are too long?

@roberth I think only subsequent warnings take 10 seconds. The first "busy" warning is printed immediately.

@roberth I think only subsequent warnings take 10 seconds. The first "busy" warning is printed immediately.

Ah, we were both wrong. First warning is after 1 second, and subsequent ones happen after 10 second intervals.

Starvation for 1 second is still bad though.

We'll need more info to really solve this. A quick search doesn't reveal a way to get useful info out of the WAL, so maybe we should do our own logging to a file, and add timestamps so that we can figure out after the fact what was blocking us.
It'd be tempting to just print what's currently blocking, but that'd be too little info, and possibly too much info, for security.

Just trying to get to the bottom of it. #10319 (comment) is still a good suggestion to move this PR forward.