Filesystems · Cluster
Log-Structured Storage
Append-only layout and garbage collection.
- 01 The Log as the Filesystem: Segments, Sequential Writes, and the End of the Free-Space Map Why writing every block at the end of a log converts scattered small updates into large sequential transfers, and what replaces the allocator once nothing is ever overwritten in place.
- 02 Inode Maps and Block Indirection: Finding Data That Never Stays in Place How a log-structured filesystem locates an inode whose address changes on every write, what the map costs per lookup, and why its cache is the deciding factor in read latency.
- 03 Checkpoint Regions and Tail Recovery in an Append-Only Log How a log-structured filesystem bounds what a crash can lose, and how it reconstructs the segment summaries written after the last durable checkpoint by scanning forward instead of backward.
- 04 Segment Liveness Accounting: How the Cleaner Knows What Is Already Dead Maintaining a live-byte count per segment as blocks are superseded, why the count must be exact rather than sampled, and how indirect blocks make the decrement easy to get wrong.
- 05 Choosing What to Clean: Greedy, Cost-Benefit, and Age-Ordered Segment Policies The three classical segment selection rules, the objective function each one optimizes, and why weighting freed space by the age of the data beats always cleaning the emptiest segment.
- 06 Decomposing Write Amplification in Log-Structured Filesystems The four terms that add up to bytes written per byte of user data, namely cleaning copies, checkpoint rewrites, summary overhead, and segment slack, and how each responds to segment size.
- 07 The Circular Log: Head, Tail, and the Distance Between Them How an append-only log wraps around a finite volume, the segment state machine that makes wrapping safe, and why the gap between the write head and the reclaim frontier sets the achievable throughput.
- 08 Garbage Collection Under Space Pressure: When Cleaning Cannot Outrun Writes What happens when a log-structured volume fills to the point that the collector must run in the foreground, why the reclaim rate collapses near full, and the reserved-capacity arithmetic that prevents it.
- 09 Data Temperature and the Cost of Cleaning Mixed Segments Why placing hot and cold data in the same segment forces the collector to copy live cold bytes to reclaim dead hot bytes, and how write segregation by expected lifetime bounds that cost.
- 10 Cleaner Interference and Foreground Latency in Append-Only Storage How a background collector steals device queue slots from the write path, why the resulting latency spikes are bursty rather than average, and the throttling and scheduling measures that bound them.