Message-ID: <gh-pgjdbc-pgjdbc-3062-c4449647421@github.com>
From: "davecramer (@davecramer)" <noreply+davecramer@github.com>
To: "pgjdbc/pgjdbc" <noreply+pgjdbc-pgjdbc@github.com>
Date: Thu, 14 May 2026 09:57:13 +0000
Subject: Re: [pgjdbc/pgjdbc] PR #3062: feat: type cache rework, codec API, and composite-type round-trip
In-Reply-To: <gh-pgjdbc-pgjdbc-3062@github.com>
References: <gh-pgjdbc-pgjdbc-3062@github.com>
List-Id: <gh-pgjdbc-pgjdbc.github.com>
X-GitHub-Author-Login: davecramer
X-GitHub-Comment-Id: 4449647421
X-GitHub-Comment-Type: issue_comment
X-GitHub-Issue: 3062
X-GitHub-Repo: pgjdbc/pgjdbc
X-GitHub-Type: comment
X-GitHub-Url: https://github.com/pgjdbc/pgjdbc/pull/3062#issuecomment-4449647421
Content-Type: text/plain; charset=utf-8

why does this require Caffeine
   
  The PR uses Caffeine as the backing cache for TypeInfoCache (OID→PgType and name→PgType lookups) and CodecRegistry
   (OID→Codec lookups). The rationale from the design spec:
  
  1. Size-based eviction — The type cache can grow unbounded if a connection encounters many user-defined types
  (composites, enums, domains). Caffeine provides LRU eviction without hand-rolling one.
  2. Concurrent access — Caffeine's Cache is thread-safe without external synchronization, which matters because
  TypeInfoCache is shared per-connection and the codec lookup happens on every getObject()/setObject() call.
  3. Avoids ConcurrentHashMap.computeIfAbsent deadlock — The spec explicitly calls out that computeIfAbsent can
  deadlock if the mapping function accesses the same map (which happens when resolving a composite type triggers
  resolution of its field types). Caffeine's get(key, loader) doesn't have this problem.
  
  Could it be done without Caffeine? Yes — a plain ConcurrentHashMap with the get() + putIfAbsent() pattern (which the
  spec already uses for fieldsByOid) would work. You'd lose automatic size-based eviction, but for a per-connection
  cache that gets cleared on epoch bump anyway, unbounded growth is unlikely to be a real problem in practice. The
  epoch-based invalidation already clears everything on DDL.
  
  The tradeoff is ~300KB added to the shaded JAR (Caffeine 2.9.3) for what is essentially a convenience over manual
  ConcurrentHashMap management. Whether that's worth it depends on how much you value the eviction policy vs. JAR
  size. For most server-side applications it's negligible; for Lambda cold starts or Android it's noticeable.