skip to content

Using the Chrome DevTools Network panel, how do you narrow down why an API call is slow for some users but fast on your machine?

level: seniorimportance: should knowfreq 46%

answer

  1. get the affected user's evidence first
  2. compare the same request, phase by phase
  3. which layer differs
  4. throttle one URL, not the whole page
  5. a local simulation is only a model

basics

~20 s

Get a sanitized HAR from an affected user, import it, and compare that call's Timing tab with yours: the phase that differs names the layer. Then test the theory locally with throttling, knowing it only approximates their network.

solid answer

~50 s

I start from evidence rather than a guess: I ask an affected user for a sanitized HAR, open it with **Import HAR file…**, and compare the same call's **Timing** tab against mine. The phase that differs names the layer. Long **Queueing** or **Stalled** suggests a client-side crowd or an HTTP/1.1 path, so I check the Protocol column. Long **DNS lookup**, **Initial connection** or **SSL** suggests a cold connection or distance. Long **Waiting for server response** suggests backend or region, and `Server-Timing` can split it. Long **Content download** suggests payload size, so I compare Size and Payload, because a heavier account may simply get more data. Then I test the theory with the **Throttling** menu, or **Throttle request URL** for that endpoint alone. Throttling is an approximation, so it supports a theory rather than proving it.

go deeper

for a junior

Know that one fast machine proves little, that a HAR export captures another person's requests, and that the Timing tab has several phases you can compare between sessions.

for a middle

Explain what each phase blames and how to compare two runs, including the Protocol, Connection ID and Size columns, so a difference in one phase turns into a specific theory.

for a senior

Lead with evidence from the affected user, separate network from data-volume causes, isolate one endpoint with a URL rule, and be honest about what throttling cannot reproduce. Finish with a claim someone can act on.

for a principal

Decide what evidence the team collects by default: field timing and Server-Timing on key APIs mean fewer requests to users for HAR files, and a cheaper answer the next time a region complains.

## Why your machine tells you little "Slow for some users" means the variable is something your setup does not have: their distance from the server, their network, their proxy, the size of their account, or how many other requests their session is making. Reloading on a fast laptop on the office network measures none of those. So the first step is to get evidence from a user who actually has the problem, and the Chrome DevTools Network panel is built to read it. ## Step 1: get an affected user's evidence Ask the user or a support colleague to open DevTools, reproduce the slowness, and click the export button in the Network panel, which writes a sanitized HAR by default. It drops the `Cookie`, `Set-Cookie` and `Authorization` headers, but URLs and bodies remain, so treat the file with care and ask for a filtered export where you can. Open it with **Import HAR file…**, or drop it onto the request list. The panel shows the imported requests with their timing and initiators, so you can read them as you would your own. ## Step 2: compare the same request, phase by phase Open the same call in both sessions and read their **Timing** tabs side by side. The total alone tells you nothing, because the phase that differs is what identifies the layer. | Phase that is long for them | What it suggests | What to check next | |---|---|---| | **Queueing** or **Stalled** | The browser held the request back | The **Protocol** column, since the six-connection limit applies only to HTTP/1.x; the number of parallel requests in their session; the **Priority** column | | **Proxy negotiation** | A proxy on their path | Whether they are on a managed network | | **DNS lookup**, **Initial connection**, **SSL** | Opening a connection is expensive for them | Distance from the server, whether their connection is reused (the **Connection ID** column) | | **Waiting for server response** | Backend work, or distance | Their region and account; the **Server Timing** section | | **Content download** | Large body or slow link | The **Size** column compared with yours | ## Step 3: compare the request itself Timing is half the comparison. The other half is whether it was really the same request. - **Payload and URL.** A different page size, filter or tenant parameter means the backend was asked for different work. - **Size.** A much larger response for one account often explains a long download and a long server wait at once, with no network problem at all. - **Neighbouring rows.** The **Waterfall** column shows what else was happening. Twenty concurrent calls from a dashboard behave differently from one call on an empty test account. ## Step 4: test the theory locally Now you have a hypothesis, try to reproduce it. 1. For a network theory, choose a preset or a custom profile from the **Throttling** menu in the Network panel. A warning icon on the panel's tab reminds you that it is on. 2. For a single endpoint, right-click its row and use **Throttle requests**, then **Throttle request URL** or **Throttle request domain**. This adds a rule to the **Request conditions** drawer, so the rest of the page keeps its normal speed. The Timing tab marks a request slowed this way with `Request was throttled`. 3. For a data theory, test with an account that has the same volume of data as the affected user. ## What throttling does not tell you Treat throttling as an approximation of a network, not as a copy of the user's network. These limits change how far you can trust the result: - **It happens inside the browser.** You choose speed figures, and the request still travels your route to your nearest server, not the user's. - **It covers the network only.** The CPU keeps your laptop's speed, so it will not reproduce a slow device handling the response. - **It cannot make a healthy backend slow.** Region-specific or account-specific server latency will not appear, however slow the profile. Reproducing the same phase profile locally is good support for a theory. Failing to reproduce it does not disprove the theory, since the user's real path remains untested. ## Reporting the result Finish with a claim someone can act on, naming the phase, the size of the gap and the evidence: "for users in that region, Initial connection plus SSL adds 700 ms on the first call, and Waiting is unchanged". If `Waiting for server response` is the phase that differs, the Network panel has done its job and the next step moves to server-side measurement, since the browser cannot see inside the server unless the server reports its own steps.

  • An affected user's HAR shows the same Timing as yours, but they still say it feels slow. What now?
    The network call is probably not the slow part. Look at how long after the response the UI updates, which is main-thread work rather than a Network panel question, or ask them to confirm which action feels slow. Also check that the export includes the right call, since an active filter or an unfinished request would leave it out.
  • Why might narrowing throttling to one URL give a better test than a slow preset for the whole page?
    A preset slows everything, which delays scripts and every other call, so a slow endpoint gets mixed up with a slow page. A rule on one URL isolates the variable and shows how the UI behaves while only that call lags, including its spinner and any race against other responses. It is also quicker to run repeatedly.
  • Only one account is slow, in a HAR from a fast office network. Which phases would you expect to differ?
    Mostly Waiting for server response, because the backend does more work for that account, and possibly Content download if the response is larger. Connection phases should match, since the network is the same. If Size is far larger for that account, the fix is more likely pagination or a query change than anything about the network.

  • Reloads on a fast laptop and concludes that nothing is wrong.
  • Compares only the total Time value and never reads the phases.
  • Treats a throttled local run that looks fine as proof the network is innocent.
  • Believes the Network panel's throttling also slows the CPU to a mobile device's speed.
  • Asks a user to send a raw HAR with sensitive data when a sanitized one would do.
  • Assumes every long Waiting for server response is the network, without checking account data size.