Glossary

76 terms, in the sense this mechanism uses them. The platform glossary has every term across all mechanisms.

Amortization: Spreading a fixed cost over many units so the cost per unit falls.Paying a fixed cost once and spreading it over many units. One delivery truck for 1,000 parcels costs less per parcel than 1,000 bicycles.
Amplification: Extra bytes read, written or stored compared with what was logically needed.If you ask for 200 bytes and the system must move 4,096 bytes, the amplification is about 20. Read amplification means extra bytes read. Write amplification means extra bytes written. Space amplification means extra bytes stored. It is the hidden cost behind most slow systems.
Anomaly detection (on spend): Alerting on deviation from a learned baseline instead of a fixed threshold.
Asymmetric scaling: Scaling up fast and down slowly; hysteresis bought on purpose to avoid flapping.
Attribution: Assigning consumption to the team, pipeline or tenant that caused it.
Auto-suspend timeout: How long a warehouse stays idle before it stops; a hysteresis parameter whose cost-optimal value equals the cold-start duration.
Autoscaling: A control loop that adds or removes capacity; it has a delay L = detect + decide + provision + warm up.
Availability zone (AZ): A separate datacenter group inside a region.
Billed = ceil(q ÷ g) × g × r: The general price formula: quantity q, granularity g, rate r.
Break-even: The point at which two options cost the same.
Cache invalidation: Knowing when a cached result is stale; an unpriced correctness cost.
Chargeback: Billing teams for their own usage; can cause teams to build private copies of shared assets.
Checkpoint: A saved snapshot of all operator state plus source offsets.1 other meaning →A saved snapshot of a job's state plus its position in the input, written to durable storage (storage that survives a crash). After a crash the job loads the last checkpoint and continues from there.
Cold start: The billed time to start a machine, runtime and caches before useful work.
Columnar files: File formats such as Parquet that store each column separately, so a query reading 5 of 200 columns reads only those 5, and that keep statistics engines use to skip parts a query cannot need.A columnar file format (such as Parquet) stores each column of a table separately, so a query that needs 5 of 200 columns can read only those 5. Partitioning splits a table into folders (for example, one per day). Pruning (or skipping) means not reading the parts a query cannot need. Selectivity is the fraction of data a query actually needs.
Commitment discount: A discount (30–60%) for promising a usage level for 1–3 years.
Compaction: Rewriting many small files into fewer larger ones.An LSM-tree (Log-Structured Merge tree) is a storage structure that never changes files in place. It collects writes in memory, writes them as sorted, immutable files, and later compacts (merges and rewrites) the files in the background to remove old versions and deleted entries. A tombstone is a marker that says "this key was deleted"; compaction removes the key and the marker. Compaction costs disk and CPU. Write amplification is the ratio of bytes the storage actually writes to bytes the application asked to write. A Bloom filter is a small structure that answers "definitely not in this file" or "maybe in this file", so reads can skip most files.
Control loop: Sense, decide, act; cannot regulate anything faster than its own delay.A cycle of sense, decide, act. A thermostat senses temperature, decides, and switches the heater. A control loop with a long delay between sensing and acting cannot regulate anything that changes faster than that delay.
Control plane: The layer that sends commands and receives status; it moves no data.The data plane is where the real data moves (storage, Spark, the warehouse). The control plane is the layer that only sends commands and receives status. The orchestrator is a control plane.
Cost-aware eviction (GDSF): Eviction by value density (frequency x miss cost / size) instead of recency only.
Cross-AZ charge: A per-byte fee for data moving between availability zones.
Data gravity: The tendency of data to attract compute because moving it is expensive.
Data skipping: Avoiding reads of blocks that statistics prove cannot match.
Dry run: Asking an engine for the estimated bytes of a query without running it.
Duty cycle (d): The fraction of time a system is active.
Effective price: List rate ÷ utilization.
Egress / ingress: Data leaving (charged) or entering (usually free) a cloud or boundary.
Elasticity: Ability to change how much capacity you own as demand changes; a trade of demand risk.
Event meter: A meter charged per request or per TB scanned.
Federation: One engine querying data in another system without copying it first.1 other meaning →
Flow meter: A meter charged per GB transferred.
Gang effect: Survival of an AND-coupled job on n nodes is (1 − p)ⁿ.
Governor: A mechanism that limits or stops expensive work.
Granularity: The smallest increment a mechanism handles or bills.The smallest piece a mechanism handles. If a courier charges per whole kilogram, one kilogram is the granularity, even if you ship 100 grams. When what you need is smaller than the granularity, you pay for the difference. This mechanism calls that difference granularity waste.
Gray failure: A partial failure that a component's own health check does not see but that clients experience.
Headroom: Spare capacity held to absorb growth during the autoscaling loop delay (ramp rate × delay).
Heavy-tailed distribution: Most items are small and a few are enormous; the top 1% carries most of the total.A distribution where a few items are enormously larger than the typical one. Query cost is an example: most queries cost almost nothing, and a few cost thousands of times more. For such data the average tells you little, and the top 1% carries most of the total.
Hysteresis: Behavior that depends on history: the threshold to enter a state differs from the threshold to leave it.A system whose behavior depends on its recent past, so it does not flip back and forth at a single threshold. A thermostat that turns on at 19 °C and off at 21 °C has a 2-degree hysteresis band. It avoids flapping, at the price of a wider swing.
Idle capacity: Capacity paid for but unused; partly mandatory (queueing floor, burstiness, granularity).
Incremental refresh: Recomputing only the changed part of a result.
Infrequent-access tier: A cheaper storage class with retrieval charges, minimum object size and minimum duration.
Interior optimum: A best value strictly between the extremes, caused by two opposed meters on one parameter.
Latency: Time one operation takes from request to response.Throughput is work finished per second. Latency is how long one operation takes. Little's Law links them to the number of operations in progress: L = λ × W, where λ is the arrival rate (per second), W is the average time one operation spends in the system, and L is how many are inside the system at the same time. Example: 1,000 requests per second, each taking 20 ms (0.02 s), means 1,000 × 0.02 = 20 requests in flight at any moment. This is an identity, not a design choice.
LIST: An object-storage call that returns about 1,000 keys per request and must be called sequentially.
Little's Law: L = λW: average items in a system = arrival rate × average time in the system.Throughput is work finished per second. Latency is how long one operation takes. Little's Law links them to the number of operations in progress: L = λ × W, where λ is the arrival rate (per second), W is the average time one operation spends in the system, and L is how many are inside the system at the same time. Example: 1,000 requests per second, each taking 20 ms (0.02 s), means 1,000 × 0.02 = 20 requests in flight at any moment. This is an identity, not a design choice.
Manifest: A small file that lists other files, with sizes, checksums and totals; it defines the set.A small file that lists other files (names, sizes, checksums). It says "this set of files is the dataset".
Materialized view: A stored query result that the system refreshes, with its admission declared.A query result that is saved as a table and kept up to date, instead of being recomputed each time it is read.
Meter: The quantity a vendor counts and bills.
Metering definition: The quantity, rounding rule, minimum and inclusion rule of a price.
Minimum billable duration / size: A floor that bills you for more than you used, to recover the vendor's fixed cost.
Multiplexing: Sharing capacity between independent users so peaks do not coincide.1 other meaning →
Object storage: A cloud service holding named objects fetched by HTTP, replaced whole, never edited.A cloud service where each file ("object") has a name and is fetched with an HTTP request. A request costs 20-100 ms before the first byte arrives, and the provider charges a small fee per request. You cannot change a few bytes inside an object. You replace the whole object.
Occupancy meter: A meter charged per machine-time, regardless of work done.
On-demand: Pay-by-the-hour machines without a promise; 2–5× the committed rate.
Outer / inner loop: The slow, accurate cost loop (attribution, forecasting) and the fast, mechanical one (caps, estimates).
P50: Sort values from small to large. P10 is the value below which 10% of them fall. P50 is the median.
Per-prefix limit: A request-rate limit tied to the key range, which depends on key distribution.
Physical data independence: The idea that a query states what it wants, not how data is stored; scan pricing makes it leak in dollars.
Pruning: Skipping whole partitions or files using metadata.A columnar file format (such as Parquet) stores each column of a table separately, so a query that needs 5 of 200 columns can read only those 5. Partitioning splits a table into folders (for example, one per day). Pruning (or skipping) means not reading the parts a query cannot need. Selectivity is the fraction of data a query actually needs.
Queue: A line of work waiting for its turn, like people waiting at a bank counter. A scheduler is the part of a system that decides who is served next.1 other meaning →
Region: A geographic cloud location.
Result cache: Stored results of earlier queries, reused when the same query recurs and inputs have not changed.
Retention: How long a log keeps records before deleting them.
Scale-to-zero: Releasing all capacity when idle; wins only when d < 1 ÷ k.
Serverless: Running each task as a short-lived function started by the platform, with no direct worker-to-worker network.
Spot: The vendor's spare machines, 60–90% below on-demand, with short-notice interruption.
Stock meter: A meter charged per GB-month.
Storage / compute separation: Keeping data in remote storage so compute can scale or stop independently.
Tagging: Labels carried in-band with resources so spend can be attributed; trace context for money.
Throughput: Work completed per unit of time.Throughput is work finished per second. Latency is how long one operation takes. Little's Law links them to the number of operations in progress: L = λ × W, where λ is the arrival rate (per second), W is the average time one operation spends in the system, and L is how many are inside the system at the same time. Example: 1,000 requests per second, each taking 20 ms (0.02 s), means 1,000 × 0.02 = 20 requests in flight at any moment. This is an identity, not a design choice.
Tier: A storage class with different rate, retrieval charge and minimums; a contract, not a medium.
Unit economics: Cost per unit of value, such as cost per active user; the trend matters more than the level.
Utilization (ρ): The fraction of time a resource is busy.Utilization (ρ, "rho") is the fraction of time a server is busy. For random arrivals, average response time grows like S ÷ (1 − ρ), where S is the service time: 2×S at 50%, 5×S at 80%, 10×S at 90%, 100×S at 99%. Queues are caused by variability meeting load, and the curve has a wall at 100%. The last few percent of utilization cost more than the first eighty.
Variability: Variation in arrival and service times. Without it a server could run near 100% busy with no queue; with it, queues form, and response time S / (1 − ρ) explodes as utilization ρ nears 1.A queue is a line of work waiting for a server. A server at 99% utilization has no line only if arrivals and service times are perfectly regular. When they vary (and they always do), queues appear. For one server with random arrivals, the average response time is R = S ÷ (1 − ρ), where S is the average service time. At ρ = 0.5 this gives 2 × S; at ρ = 0.9 it gives 10 × S; at ρ = 0.99 it gives 100 × S. The last few percent of utilization are very expensive in waiting time.
Warm pool: Pre-started spare instances that reduce provisioning delay.
Workload routing: Sending queries to the engine whose cost curve is cheapest for them.