Evm_node_lib_dev.Sqlite_compression_migrationOne-shot startup migration that zstd-compresses the legacy uncompressed rows of the target BLOB columns listed below, then reclaims disk space via VACUUM INTO followed by a file rename. Interruptible: progress is persisted in a lazily-created compression_migration cursor table (one row per target table plus a __vacuum__ sentinel for the vacuum + rename step), so a crash or restart resumes where the previous run left off.
The migrated columns are:
blueprints.payloadblocks.blocktransactions.receipt_fieldstransactions.object_fieldsblocks.tez_block is also wrapped in zstd_compress / zstd_decompress on the write/read path in Evm_store.Q, but is intentionally excluded from the migration scan: the column is unused — and therefore empty — on every production EVM node, so a row scan would be wasted I/O. New writes go through zstd_compress regardless, so any future use of the column starts compressed.
On steady-state startups (every cursor row has done_ = 1), the function exits in O(1) without touching any target table.
On VACUUM INTO / rename failure, the cursor table is preserved with done_ = 1 for every fully compressed table and done_ = 0 on the __vacuum__ sentinel — so the next startup retries only the vacuum + rename phase, never the multi-hour row scan. The error is propagated to the caller.
A narrow crash window exists between Unix.rename succeeding and the __vacuum__ sentinel being marked done_ = 1 (essentially the time it takes to reopen the store and run a single UPDATE). A crash in that window leaves the file fully migrated and vacuumed but with the sentinel still at 0; the next run skips every table loop (each target is already done_ = 1) and re-runs VACUUM INTO once on the already-compressed file. That second vacuum is idempotent — output is identical — but costs the same wall time as the first. We accept the cost: closing the window would require writing the sentinel into the tmp DB before the rename, which is more machinery than the rarity of the scenario justifies.
val run_if_needed :
store:Octez_sqlite.Sqlite.t ->
path:string ->
open_store:
(unit -> Octez_sqlite.Sqlite.t Tezos_base.TzPervasives.tzresult Lwt.t) ->
Evm_node_config.Configuration.sqlite_compression_config ->
Octez_sqlite.Sqlite.t Tezos_base.TzPervasives.tzresult Lwt.trun_if_needed ~store ~path ~open_store cfg runs the compression migration and returns the (possibly reopened) store. ~open_store is used to reopen after VACUUM INTO + rename. Must be called after the store has been opened and before any client work runs against it.