CDRIVER-3645 Close operation connections to removed servers #1618
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Track last known server IDs in the client pool. Before doing an expensive prune, do a cheaper check to see if server IDs changed.
Servers may be removed from the topology in other scenarios aside from SRV polling (e.g. during SDAM)
eramongodb
requested changes
May 28, 2024
Co-authored-by: Ezra Chung <[email protected]>
Reduces the scope holding the shared topology lock.
eramongodb
approved these changes
May 28, 2024
vector-of-bool
approved these changes
May 29, 2024
Comment on lines
384
to
395
| static bool | ||
| maybe_prune (void *item, void *ctx_) | ||
| { | ||
| mongoc_cluster_node_t *cn = (mongoc_cluster_node_t *) item; | ||
| prune_ctx *ctx = (prune_ctx *) ctx_; | ||
| uint32_t server_id = cn->handshake_sd->id; | ||
|
|
||
| if (!bsearch (&server_id, ctx->server_ids->data, ctx->server_ids->len, sizeof (uint32_t), server_id_cmp)) { | ||
| mongoc_cluster_disconnect_node (ctx->cluster, server_id); | ||
| } | ||
| return true; | ||
| } |
There was a problem hiding this comment.
Recommend adding commentary on what/why this function is does what it does.
Comment on lines
410
to
430
| for (; idx1 < cluster->nodes->items_len; idx1++) { | ||
| // Compare both sorted lists in order. `cluster->nodes` may be a smaller list if not all servers were used. | ||
| mongoc_set_item_t *cn = &cluster->nodes->items[idx1]; | ||
| bool found = false; | ||
| for (; idx2 < known_server_ids->len; idx2++) { | ||
| uint32_t last_known = _mongoc_array_index (known_server_ids, uint32_t, idx2); | ||
| if (cn->id == last_known) { | ||
| found = true; | ||
| break; | ||
| } | ||
| } | ||
| if (!found) { | ||
| // A server in the cluster is not in the last known server ids. Prune it. | ||
| needs_prune = true; | ||
| break; | ||
| } | ||
| } | ||
|
|
||
| if (!needs_prune) { | ||
| return; | ||
| } |
There was a problem hiding this comment.
Is it likely that this early-return saves a significant amount of time for the additional complexity that it adds?
There was a problem hiding this comment.
Good question. Another benchmark does not show significant change with this early-return removed:
| Test | Baseline (all conditional pruning) | Without conditional client pruning | Without conditional pruning (simple prune) |
|---|---|---|---|
| Parallel/Pool/ShareClients/Threads:1 | 249159 | 246554 (-1.05%) | 235143 (-5.63%) |
| Parallel/Pool/ShareClients/Threads:10 | 1312571 | 1313206 (0.05%) | 866941 (-33.95%) |
| Parallel/Pool/ShareClients/Threads:100 | 1932369 | 1917569 (-0.77%) | 1194404 (-38.19%) |
| Parallel/Pool/ShareClients/Threads:150 | 1555300 | 1564356 (0.58%) | 1564161 (0.57%) |
The "Baseline" is the current changeset. The "Without conditional client pruning" removes this early-return.
Updated to remove the early-return.
kkloberdanz
approved these changes
May 30, 2024
And rename `server_ids` to `known_server_ids` for consistency
kevinAlbs
added a commit
to kevinAlbs/mongo-c-driver
that referenced
this pull request
May 30, 2024
…1618) * add regression test: check connection counts of servers removed by SRV polling * prune clients on push Track last known server IDs in the client pool. Before doing an expensive prune, do a cheaper check to see if server IDs changed. * test direct topology changes Servers may be removed from the topology in other scenarios aside from SRV polling (e.g. during SDAM) * document expected use of pop/push --------- Co-authored-by: Ezra Chung <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Close operation connections to removed servers.
Verified with this patch build: https://spruce.mongodb.com/version/665489593f686b000739adf0
Background & Motivation
Motivated by an issue reported in HELP-59749. Tests observed an unexpected increasing connection count. This may have been triggered by (misconfigured?) DNS records resulting in servers being frequently added/removed from the topology during SRV polling.
Rejected alternative: callback
I considered adding a callback to
_mongoc_topology_description_remove_serverto more immediately signal to close connections. This was rejected to reduce risk of an AB-BA deadlock._mongoc_topology_description_remove_serverexpects the caller to holdmongoc_topology_t::tpld_modification_mtx. Modifying pooled client requires holdingmongoc_client_pool_t::mutex.There is an existing pattern of locking
mongoc_client_pool_t::mutexthenmongoc_topology_t::tpld_modification_mtx(viamongoc_client_pool_try_pop=>_start_scanner_if_needed=>_mongoc_topology_background_monitoring_start=>mc_tpld_modify_begin).Instead, this PR proposes lazily pruning stale connections on
mongoc_client_pool_push. I expect multi-threaded applications to pop/push clients frequently to encourage reuse of clients among threads. A note is added to documentation to make this recommendation explicit.Performance Impact
A benchmark was run to compare throughput of repeated use of the pool through pop+command+push with pruning:
An initial solution ("With simple pruning") iterated over all pooled clients to prune on each call to
mongoc_client_pool_push. Due to the performance impact, checks were added to only prune if topology description change was detected ("With conditional pruning"). This motivated storing thelast_known_serveridsin themongoc_client_pool_tto quickly determine if the set of server IDs changes. I expect the set of server IDs changing is a "rare" event (e.g. HELP-59749 notes an extreme case where SRV records unexpectedly change. Even then, I expect the set of server IDs would change about once per minute).