Hudi — Middle¶

At middle level, focus on this question:

When an update arrives, what actually happens to the underlying files under Copy-on-Write versus Merge-on-Read?

Prerequisite: junior.md.

Copy-on-Write (COW): rewrite the affected file immediately¶

flowchart LR Update["UPDATE customer_id=42"] --> Locate["Locate the file\ncontaining customer 42's\ncurrent row"] Locate --> Rewrite["Rewrite that ENTIRE file\n(even though only 1 row\nchanged) with the\nupdated row included"] Rewrite --> NewFile["New version of the file,\nold version superseded"]

Under COW, every update triggers a full rewrite of whichever data file contains the affected row(s) — even if only one row out of 100,000 in that file changed, the entire file is rewritten. Reads are simple and fast (just read the current files directly, no merge logic needed at read time), but writes are expensive under high update volume, since each update potentially rewrites a large file.

Merge-on-Read (MOR): log the change, merge lazily at read time¶

flowchart LR Update["UPDATE customer_id=42"] --> Log["Append the change to a\nsmall DELTA LOG file -\nNO base file rewrite"] Log --> ReadTime["At READ time: merge the\nbase file + delta log\nON THE FLY to produce\nthe current view"]

Under MOR, an update is appended cheaply to a small delta/log file (structurally similar to the LSM-tree's memtable-then-flush pattern from the LSM-Tree professional page) — the expensive base-file rewrite is deferred, happening only during periodic compaction. Reads must merge the base file with any pending delta log entries on the fly, which costs extra read-time work, but writes are dramatically cheaper for high-update-volume workloads.

	Copy-on-Write	Merge-on-Read
Write cost	High (full file rewrite per update)	Low (cheap log append)
Read cost	Low (direct file read)	Higher (merge base + log at read time)
Fits	Read-heavy, moderate update volume	Write-heavy, high update volume

🎓 Takeaway: COW and MOR are the exact same "pay now vs. pay later" trade-off as the LSM-tree's compaction-strategy choice and Delta Lake's deletion-vector-versus-immediate-rewrite decision — Hudi just makes this a first-class, explicit table type choice rather than an implementation detail.

Test yourself¶

Why does COW's write cost scale with file size, even for a single-row update?
Why does MOR's read cost include merge work that COW's read path doesn't need at all?
For a table receiving 100,000 small updates per hour, which table type would you choose, and why?

Continue to senior.md.