Struct mz_persist_client::cfg::DynamicConfig

pub struct DynamicConfig {
    batch_builder_max_outstanding_parts: AtomicUsize,
    compaction_heuristic_min_inputs: AtomicUsize,
    compaction_heuristic_min_parts: AtomicUsize,
    compaction_heuristic_min_updates: AtomicUsize,
    compaction_memory_bound_bytes: AtomicUsize,
    gc_blob_delete_concurrency_limit: AtomicUsize,
    state_versions_recent_live_diffs_limit: AtomicUsize,
    usage_state_fetch_concurrency_limit: AtomicUsize,
}

Persist configurations that can be dynamically updated.

Persist is expected to react to each of these such that updating the value returned by the function takes effect in persist (i.e. no caching it). This should happen “as promptly as reasonably possible” where that’s defined by the tradeoffs of complexity vs promptness. For example, we might use a consistent version of BLOB_TARGET_SIZE for the entirety of a single compaction call. However, it should never require a process restart for an update of these to take effect.

These are hooked up to LaunchDarkly. Specifically, LaunchDarkly configs are serialized into a mz_dyncfg::ConfigUpdates. In environmentd, these are applied directly via mz_dyncfg::ConfigUpdates::apply to the PersistConfig in crate::cache::PersistClientCache. There is one PersistClientCache per process, and every PersistConfig shares the same Arc<DynamicConfig>, so this affects all DynamicConfig usage in the process. The ConfigUpdates is also sent via the compute and storage command streams, which then apply it to all computed/storaged/clusterd processes as well.

Fields

batch_builder_max_outstanding_parts: AtomicUsize

compaction_heuristic_min_inputs: AtomicUsize

compaction_heuristic_min_parts: AtomicUsize

compaction_heuristic_min_updates: AtomicUsize

compaction_memory_bound_bytes: AtomicUsize

gc_blob_delete_concurrency_limit: AtomicUsize

state_versions_recent_live_diffs_limit: AtomicUsize

usage_state_fetch_concurrency_limit: AtomicUsize

Implementations

impl DynamicConfig

const LOAD_ORDERING: Ordering = Ordering::SeqCst

const STORE_ORDERING: Ordering = Ordering::SeqCst

pub fn batch_builder_max_outstanding_parts(&self) -> usize

The maximum number of parts (s3 blobs) that crate::batch::BatchBuilder will pipeline before back-pressuring crate::batch::BatchBuilder::add calls on previous ones finishing.

pub fn compaction_heuristic_min_inputs(&self) -> usize

In Compactor::compact_and_apply, we do the compaction (don’t skip it) if the number of inputs is at least this many. Compaction is performed if any of the heuristic criteria are met (they are OR’d).

pub fn compaction_heuristic_min_parts(&self) -> usize

In Compactor::compact_and_apply, we do the compaction (don’t skip it) if the number of batch parts is at least this many. Compaction is performed if any of the heuristic criteria are met (they are OR’d).

pub fn compaction_heuristic_min_updates(&self) -> usize

In Compactor::compact_and_apply, we do the compaction (don’t skip it) if the number of updates is at least this many. Compaction is performed if any of the heuristic criteria are met (they are OR’d).

pub fn compaction_memory_bound_bytes(&self) -> usize

The upper bound on compaction’s memory consumption. The value must be at least 4*blob_target_size. Increasing this value beyond the minimum allows compaction to merge together more runs at once, providing greater consolidation of updates, at the cost of greater memory usage.

pub fn gc_blob_delete_concurrency_limit(&self) -> usize

The maximum number of concurrent blob deletes during garbage collection.

pub fn state_versions_recent_live_diffs_limit(&self) -> usize

The # of diffs to initially scan when fetching the latest consensus state, to determine which requests go down the fast vs slow path. Should be large enough to fetch all live diffs in the steady-state, and small enough to query Consensus at high volume. Steady-state usage should accommodate readers that require seqno-holds for reasonable amounts of time, which to start we say is 10s of minutes.

This value ought to be defined in terms of NEED_ROLLUP_THRESHOLD to approximate when we expect rollups to be written and therefore when old states will be truncated by GC.

pub fn usage_state_fetch_concurrency_limit(&self) -> usize

The maximum number of concurrent state fetches during usage computation.

pub fn set_compaction_memory_bound_bytes(&self, val: usize)

Trait Implementations

impl Debug for DynamicConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.