activerecord-virgodb-adapter 0.1.4 → 0.1.6

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d6376b065e29146c096491b6ce2a68e4331e806161a570b6549b068561a85f1f
-  data.tar.gz: 1555e75fdf0c2418c1d3439fb1940f7dd284e48e28109248ffd249454b41dc55
+  metadata.gz: '0390e08c3ff3f57f95bce4609600461132e9af55122f1d9be076bd59171bc37d'
+  data.tar.gz: 216fb2576dd924f68d9e0e635824a825b0bc6863db6f3694438943343a630df3
 SHA512:
-  metadata.gz: ab5ef2e0113e6c38295495cb10d1c2ca4bb77d135b65a636cf58aee8777146bc29795b1c7d3d2950898735a81367e2d88e3ce235a3d4f567b7e934ae8d69e885
-  data.tar.gz: 7375e48809a1b05cc4a1ae8add6ed5e1a36d33a4564e0c711a8a8228e910e4f79d9ad039aaeccd66b839b88a4ff31640aec9e5d0f9ac246398d320b6b1224611
+  metadata.gz: d3bfd3b9cdad22f0385a1f9d13a84483aa26c89cc3c16f38d231f8c8885b2af9a1beac0ef7d53ee6237b87092c074026fab6370fab4b605230342a81a8c574b5
+  data.tar.gz: 52c0f3df23b6a21a97a6322f58d59aa2c75511c998a21ce4e3eb54d8049f674c97d1b61061f377fcd1e028832669b7573033d2a6409804ffc241749cf91492bf

data/lib/active_record/tasks/virgodb_database_tasks.rb

@@ -72,6 +72,7 @@ module ActiveRecord
 
       def structure_dump(filename, extra_flags)
         super
+        strip_null_autoindex_blank_lines!(filename)
         reformat_create_table_statements!(filename)
         annotate_physical_layout!(filename)
         dump_schema_versions!(filename)
@@ -194,23 +195,69 @@ module ActiveRecord
         File.write(filename, content)
       end
 
-      # `SQLite3Adapter#structure_dump`'s raw `SELECT sql FROM sqlite_master`
-      # path (used whenever `ActiveRecord::SchemaDumper.ignore_tables` is
+      # `SQLiteDatabaseTasks#structure_dump`'s `ignore_tables`-triggered
+      # fallback (used whenever `ActiveRecord::SchemaDumper.ignore_tables` is
       # non-empty, e.g. excluding tables owned by a separate replication/
-      # backup mechanism) reproduces each
-      # CREATE TABLE exactly as ActiveRecord originally generated it --
-      # one long line, every column crammed together. Fine for SQLite to
-      # load back (structure_load doesn't care about whitespace), useless to
-      # a human reading the dump next to `db/clickhouse_structure.sql`
-      # (ClickHouse's own `SHOW CREATE TABLE` is naturally one-column-per-line,
-      # for comparison). Purely cosmetic, safe to do unconditionally: only
-      # rewrites the exact same statement with different whitespace, no DDL
-      # semantics change. Skips statements that already span multiple lines
-      # (e.g. this adapter's own hand-written `schema_versions` DDL, created
-      # via a real newline-containing heredoc) -- nothing to reformat there.
+      # backup mechanism -- exactly hintpot's own litestream initializer)
+      # runs `SELECT sql || ';' FROM sqlite_master WHERE tbl_name NOT IN
+      # (...) ORDER BY tbl_name, type DESC, name` through the `sqlite3` CLI.
+      # Any table with a non-INTEGER (so non-rowid-alias) PRIMARY KEY --
+      # `maintenance_lease`, `flush_completions`, `ar_internal_metadata`,
+      # `schema_migrations`, all `TEXT PRIMARY KEY` -- gets a second,
+      # *hidden* row in `sqlite_master` for its `sqlite_autoindex_*`, whose
+      # `sql` column is NULL (an autoindex has no CREATE INDEX text of its
+      # own; SQLite generates it implicitly). `NULL || ';'` is NULL, and the
+      # `sqlite3` CLI's default `.nullvalue` prints a NULL result as nothing
+      # -- so each such table silently donates one stray blank line to the
+      # dump, unrelated to any table's real position or content (confirmed
+      # directly against a real `sqlite_master`; every non-blank-line-
+      # producing table here is either INTEGER PRIMARY KEY, a rowid alias
+      # needing no separate index, or has no primary key at all). The other
+      # `structure_dump` path (`.schema --nosys`, taken when
+      # `ignore_tables` is empty) already skips these rows itself --
+      # confirmed empirically too -- so this is purely a gap in the
+      # raw-SQL fallback, not something either path does on purpose.
+      # Stripping now, before `reformat_create_table_statements!` touches
+      # anything, means later steps never have to reason about which blank
+      # lines are "real" -- there aren't any yet at this point in the
+      # pipeline (`dump_schema_versions!`'s own deliberate ones are added
+      # after this runs).
+      def strip_null_autoindex_blank_lines!(filename)
+        content = File.read(filename)
+        content.gsub!(/\n\n+/, "\n")
+        File.write(filename, content)
+      end
+
+      # Reformats one-line CREATE TABLE statements (both `structure_dump`
+      # paths -- see `strip_null_autoindex_blank_lines!` above for the two --
+      # produce these one column crammed after another) into one column per
+      # line. Fine for SQLite to load back either way (structure_load
+      # doesn't care about whitespace), useless to a human reading the dump
+      # next to `db/clickhouse_structure.sql` (ClickHouse's own `SHOW CREATE
+      # TABLE` is naturally one-column-per-line, for comparison). Purely
+      # cosmetic, safe to do unconditionally: only rewrites the exact same
+      # statement with different whitespace, no DDL semantics change.
+      #
+      # Must handle statements that ALREADY span multiple lines, not just
+      # single-line ones: every virgodb-side manifest table (`crates/
+      # manifest/src/lib.rs`'s `maintenance_lease`, `orphaned_files`, this
+      # adapter's own `schema_versions`, ...) is bootstrapped from a
+      # multi-line, already-indented `CREATE TABLE` literal, and any column
+      # added to one of those later via `ALTER TABLE ... ADD COLUMN` (the
+      # only way SQLite can add a column to an existing table) gets its
+      # definition spliced into that SAME stored multi-line text, right
+      # before the final `)` -- with none of the original literal's
+      # indentation, producing exactly the lopsided
+      # "seven clean columns, then a cramped `, col1, col2)` tail" shape a
+      # single-line-only regex would silently skip over and leave raw.
+      # `/m` (dot matches newline) plus a non-greedy `(.*?)` is required
+      # together: `/m` alone with the old greedy `(.*)` would swallow every
+      # later `CREATE TABLE`/`CREATE INDEX` statement up to the LAST `);` in
+      # the file into one match, since nothing would stop it at the nearer
+      # one.
       def reformat_create_table_statements!(filename)
         content = File.read(filename)
-        content.gsub!(/^CREATE TABLE (IF NOT EXISTS )?("[\w]+"|\w+) \((.*)\);$/) do
+        content.gsub!(/^CREATE TABLE (IF NOT EXISTS )?("[\w]+"|\w+) \((.*?)\);$/m) do
           if_not_exists = Regexp.last_match(1)
           table_name = Regexp.last_match(2)
           columns = split_top_level_commas(Regexp.last_match(3))

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: activerecord-virgodb-adapter
 version: !ruby/object:Gem::Version
-  version: 0.1.4
+  version: 0.1.6
 platform: ruby
 authors:
 - virgodb