Add list compression tuning flags (#7762)

Rename the ZSTD dictionary threshold flag to a non-experimental name and add a separate list compression level flag used when building QList dictionaries.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

chore(server): migrate main_service handlers to CmdArgParser

Migrate most main_service.cc command handlers from CmdArgList to
CmdArgParser, following the same pattern as the recent set_family.cc
and zset_family.cc migrations.

- Update Quit, Multi, Watch, Unwatch, Discard, Exec, Publish,
  Subscribe, Unsubscribe, PSubscribe, PUnsubscribe, Function,
  Monitor, Pubsub, Command handlers
- Propagate ParsedArgs through ChangeSubscription/ChangePSubscription
  in conn_context.h/.cc to avoid CmdArgList conversion
- Update MFUNC macro to call MakeParserFromContext instead of
  passing CmdArgList; keep MFUNC_OLD for Eval* handlers (deferred)

Signed-off-by: Roman Gershman <roman@dragonflydb.io>

feat(server): Concurrent squashing (#7597)

fix(facade): retire experimental_io_loop_v2 flag (#7605)

The experimental_io_loop_v2 flag was accidentally dropped.
Using ABSL_RETIRED_FLAG instead of simply removing it preserves
backward compatibility for deployments that reference it in configs
or scripts.

Signed-off-by: Gil Levkovich <69595609+glevkovich@users.noreply.github.com>

fix: oom on incoming slot migration does not clean the replica (#7533)

Signed-off-by: Kostas Kyrimis <kostas@dragonflydb.io>

fix: batched backpressure (#7515)

facade: Trim parsed args if capacity much higher than size (#7386)

* facade: Trim parsed args if overallocated

On parsing new command, if cap > 2* prev size needed, then trim command
to avoid holding onto oversized capacity. Also do this only over 256 KiB
to not churn on small commands and only correct large allocs

Signed-off-by: Abhijat Malviya <abhijat@dragonflydb.io>
(cherry picked from commit 0a369ac)

* facade: Do not use clear to drop memory

This API has changed in recent versions of absl and no longer
deallocates

Signed-off-by: Abhijat Malviya <abhijat@dragonflydb.io>
(cherry picked from commit c062d40)

---------

Signed-off-by: Abhijat Malviya <abhijat@dragonflydb.io>

fix: handle incompressible data in HuffmanCheckTask (#7251)

HuffmanCheckTask crashes the process with SIGABRT when key data is
incompressible (e.g. random hashes, UUIDs, encrypted blobs):

  F huff_coder.cc:108] Check failed: !HUF_isError(size)

The task runs as an idle task on shard 0 once key memory exceeds 50 MB.
It builds a Huffman frequency table from sampled key data, then
unconditionally calls HuffmanEncoder::Export() to serialize the table
for logging. When the data is high-entropy (compression ratio = 1.0),
HUF_writeCTable_wksp returns an error for the degenerate table and the
CHECK kills the process. Since the task re-triggers on every restart
after snapshot recovery, this results in a crash loop.

Note: --compression_mode does not prevent this. That flag controls
snapshot serialization compression (LZ4/ZSTD), not the background
HuffmanCheckTask.

Fix:
- Export() returns std::optional<std::string> instead of crashing via
  CHECK. Logs a warning on failure.
- HuffmanCheckTask::Run() skips Export() when compression ratio >= 1.0,
  since there is no useful table to log.
- All callers (debugcmd.cc, tests) updated to handle the optional.

test(connection): add a fix for FIN_WAIT_2 leak bug with BLPOP (#7224)

* test(connection): add reproduction for FIN_WAIT_2 leak with BLPOP

Closing a client socket while the connection's reader fiber is parked on
a blocking command (BLPOP) does not always wake the fiber to observe EOF.

Under sustained concurrent open+close churn, a fraction of FDs linger on
the server in CLOSE_WAIT (client in FIN_WAIT_2) until the command times
out or the client is CLIENT KILL'd, matching the reaper-driven leak
reported in production. valkey closes its half immediately
on FIN even mid-BLPOP.

The bug is fixed in helio by romange/helio#580

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Roman Gershman <roman@dragonflydb.io>