Optimization

Insights

See the latest cost anomalies, tag coverage and cost alerts, and which services and regions changed most.

Insights sits under Optimization in the product's sidebar. It brings together the newest cost anomalies still wanting attention, tag coverage, the workspace's cost alerts, the fastest-growing services and the largest cost changes. View savings opens Savings, the queue of cost that can be removed.

Before you start

  • Connect at least one source, so the workspace has cost data. See Manage connections.
  • Any team role can view Insights. Acknowledging, archiving and ignoring an anomaly, and creating, editing, pausing and deleting cost alerts, needs the team role Editor with edit access to the workspace.

The Cost anomalies card on the Insights page, listing open anomalies with their severity, description, expected and actual cost, and actions

Cost anomalies on Insights for the Production workspace.

Review anomalies

  1. Open Insights. The "Cost anomalies" card opens on "To review", the five newest of the workspace's anomalies that are open or acknowledged, and says how many are open. "Show all …" lists every one of them, and "Show more" reads the next 50. "Archived" and "Ignored" list the anomalies taken off the list, newest first. "How anomalies are detected" opens the rules detection follows. Another workspace's anomalies never appear here, and neither does one whose connection has since left the workspace.

  2. Each anomaly names the service, its severity (low, medium, high or critical), a description, the "Expected" and "Actual" cost, how far the actual cost rose, the day it was charged and when it was detected. Under "Top contributors to the increase" it lists up to three resources that added most to the rise, each with its service, account and amount, and when one resource accounts for more than half of the rise it says "Mostly from" that resource. When one region accounts for more than half of the rise, the region shows beside the service. Under "Top SKUs behind the increase" it lists up to three SKUs that added most, each with its amount, named as a cost report's SKU grouping names them: the provider's SKU label, such as an AWS usage type, or the SKU's ID when it has none. The notification for a new anomaly names the same resources, SKUs and region. New spend shows "New" in place of how far the cost rose: its expected cost is the rest of the service's cost that day, and its description says how much is new and, when one region holds more than half of it, where.

  3. "Open in report" opens a new cost report with the evidence: the anomaly's provider and service, and its region when one is named, by day from 14 days before the day it was charged to 3 days after, grouped by resource ID, or by SKU when no single resource accounts for most of the rise. The report counts every charge of its filter, so its day can be higher than the anomaly's actual cost by commitment fees or marketplace charges the workspace's anomaly settings leave out, and with two connections to one provider it shows both.

  4. Decide what it is:

    • "Acknowledge" marks it as seen. It stays under "To review", marked "Acknowledged".
    • "Archive" moves it to "Archived".
    • "Ignore" moves it to "Ignored" with a reason you give, such as a planned load test.

    Each action says what it does while you point at it. After "Archive" or "Ignore", "Undo" in the message that confirms it puts the anomaly back as it was. Costfluent records who changed the status and when, and detection never changes it again.

While an issue tracker is connected, each anomaly also offers Create issue to anyone who can change its status, and shows its issue once one is asked for.

Under "Archived" and "Ignored", each anomaly says why it is there: the reason it was ignored for, or, when it archived itself, that nobody acted on it within a week of the day it was charged or that restated cost data no longer supports it. "Reopen", for anyone who can change its status, puts it back under "To review".

With nothing to review it reads "No anomalies detected. Your costs are looking normal." The Alerts page lists every recent anomaly on the workspace's connections under What fired, where "Open in report" and "Acknowledge" work as they do here.

Read tag coverage and cost alerts

  • The "Tag coverage" card shows how much of the cost is on tagged resources, and "Set alert" creates a coverage alert. See Tags and tag coverage.
  • The Cost alerts card lists the workspace's cost alerts with their Name, Status and Threshold, with actions to view and edit, pause, resume and delete each one, and New alert creates a spend threshold, percentage change or month-to-date spend alert. See Alerts.

Read the trend cards

The trend cards and the "Tag coverage" card always cover the last 30 days, ending today. Insights has no date control of its own.

  1. Fastest growing services lists up to five services whose cost rose most against the period before, with both amounts and the change in percent. Both cover the days through the last settled day, normally two days ago, and as many days before, because a provider delivers the last day or two late; the card then reads "Services with the highest cost increase vs the same days before, through …", and Significant cost changes names the same day.
  2. Significant cost changes lists up to five regions or other dimensions whose cost changed most, and a Total period change row when the total moved. When the workspace converts currencies and some spend has no exchange rate for its days, that spend is in neither total, and a note under both cards names each period's amount. When a billing currency converts in one period and not in the other, the two totals would not cover the same spend, so the cards compare nothing, and the cards and the note say so; see Currency settings.
  3. Use View in cost reports to open Cost reports on the workspace's default report; see Explore cost data.

When nothing grew or changed, the cards say "No significant growth detected" and "No significant changes detected".

When the two periods could not be compared, both cards say so instead, and why: the costs are in several currencies with conversion off, or a currency converts in one period and not in the other.

Explain a month's change

Planned: change investigation

Not yet available in Costfluent.
  1. Open Insights and select the Change tab. It opens on the last complete calendar month.
  2. Pick one of the last twelve complete months under Month, and what to rank under Break down by: Service, Resource, Account or Owner. Owner groups cost by the owners the workspace's segments assign.
  3. Read the result. The heading gives the month's total and how much it changed against the same number of days before the month. The table lists the ten names whose cost changed most, up or down, each with Before, This month, Change and What happened: "New", "Gone", "Up", "Down" or "Unchanged". The rest are summed in one row.
  4. Select a name to open the cost rows behind it: a new cost report of its cost over the month, through today for the current month, by day, unamortized as the change is, compared with the same number of days before the month and grouped by the next dimension down (a service by SKU, a resource or an account by service, an owner by service). The summed row opens the names the table does not list, grouped as the table is. An owner opens the segments that carry its name; "Unallocated", the summed row of owners and a name the provider left empty, which reads "Not reported", open none. The report shows the workspace's currency, where the table shows each billing currency apart.
  5. To record why the cost changed, save the report and add a note to it on the day it explains. See Annotations.

With no cost in the month or the days before it, the tab reads "No cost in either period".

How anomalies are detected

  • Anomaly detection runs every four hours. It scores each service's daily cost on each of the workspace's connections, with the money settings and in the currency the workspace's reports use, for the days two to four days ago: no provider has complete data for today or yesterday. Every run scores those days again, so a late or restated figure updates the anomaly rather than adding a second one.
  • A day is an anomaly only when all of these hold:
    • its cost is more than two standard deviations above the previous thirty days, and at least seven of those days had cost;
    • it is at least 5 above that average, in the workspace's currency;
    • it is at least 0.5% of the workspace's cost that day;
    • it is more than 20% above the average of the seven days before;
    • it is above the day before;
    • it is more than 20% above the median of the same weekday in each of the four weeks before, a day without cost counting as 0, so a job that ran on that weekday in three of those weeks at the same cost is routine;
    • it is more than 20% above the same day of the previous month, or the previous month's last day when it is its month's last, so a monthly fee that posts as it did the month before is routine;
    • the same service on the same connection had no anomaly in the seven days before.
  • A service, or a region of a service, that starts costing money is an anomaly of its own, new spend, which the rules above cannot catch: they need a week of history. It is told on the first day its cost is at least 5 and at least 0.5% of the workspace's cost that day, when that day is one of the first seven since it was first billed and nothing was billed for it, in any currency, from 95 days before that day until its first. When the service itself is new, its regions are not counted again. Nothing is new spend on a connection that had no cost from 95 to 31 days before the day: on a connection added recently, every service is new to Costfluent rather than to you. A day that is both is told as a spike. An anomaly in the seven days before never holds back new spend, and a spike it holds back is told as the new spend its day holds, if any.
  • Marketplace charges (products bought from AWS Marketplace and Azure Marketplace sellers) and commitment fees (reservation and Savings Plan purchases and their recurring fees) are left out of detection: they are bought rather than used, so a jump in one is no anomaly. A workspace can include either under Anomaly detection on its settings page.
  • Only increases are anomalies. A connection whose cost falls to nothing shows that in its own health instead.
  • When the workspace does not convert currencies, a connection that bills in two currencies is scored as two series, never summed.

How it behaves

  • An open anomaly nobody acted on archives itself a week after the day it was charged. One that restated data no longer supports archives itself too, and opens again if a later figure supports it once more; so does a spike that the weekday or month rule holds back when its day is scored again.
  • A new anomaly creates an in-product notification in the workspace and goes, once, to the notification channels the workspace chose under Anomaly notifications on its settings page; with none chosen it stays in the product. An anomaly whose cost rose less than the workspace's threshold under Anomaly detection (Notify from, in the currency the workspace shows its figures in, which is its exchange rate policy's while one is in force; for new spend, the new spend) notifies no one and still shows here; with the threshold at 0, every anomaly notifies, whatever its severity. Scoring the same day again never notifies again.
  • Changing an anomaly's status does not change any cost figure.
  • Anomalies are also available through the Public API. Listing them needs the View anomalies capability and acknowledging them Acknowledge anomalies. A workspace token reaches only its own workspace's anomalies.
  • The Change tab counts billed cost in the currency the provider billed it in, with one table per currency, before any conversion or amortisation, so its figures can differ from a report's. Like a report with the default settings, it leaves out credits, refunds and corrections (see Cost amounts), so a credit does not show as a fall in a service's spend. Months are calendar months in UTC.
  • A month whose comparison period reaches back past your plan's history shows an error instead of a table.