A layout of a transaction log enables efficient logging of metadata into entries of the log, as well as efficient reclamation and recovery of the log entries by a volume layer of a storage input/output (I/O) stack executing on one or more nodes of a cluster. The transaction log is illustratively a two stage, append-only logging structure, wherein the first level is non-volatile random access memory (NVRAM) embodied as a NVlog and the second stage is disk, e.g., solid state drive (SSD). During crash recovery, the log entries are examined for consistency and scanned to identify those entries that have completed and those that are active, which require replay. The log entries are walked from oldest to newest (using sequence numbers) searching for the highest sequence number. Partially complete log entries (e.g., log entries in-progress when a crash occurs) may be discarded for failing a checksum (e.g., a CRC error). Old value/new value logs may be used to implement roll-forward or roll-back semantics to replay the log entries and fix any on-disk data structures, first from NVRAM and then from on-disk logs.